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 int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 299                 loff_t blksize)
 300 {
 301         if (!blksize)
 302                 blksize = NBD_DEF_BLKSIZE;
 303         if (blksize < 512 || blksize > PAGE_SIZE || !is_power_of_2(blksize))
 304                 return -EINVAL;
 305
 306         nbd->config->bytesize = bytesize;
 307         nbd->config->blksize = blksize;
 308
 309         if (!nbd->task_recv)
 310                 return 0;
 311
 312         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 313                 nbd->disk->queue->limits.discard_granularity = blksize;
 314                 nbd->disk->queue->limits.discard_alignment = blksize;
 315                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 316         }
 317         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 318         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 319
 320         if (max_part)
 321                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 322         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 323                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 324         return 0;
 325 }
 326
 327 static void nbd_complete_rq(struct request *req)
 328 {
 329         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 330
 331         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 332                 cmd->status ? "failed" : "done");
 333
 334         blk_mq_end_request(req, cmd->status);
 335 }
 336
 337 /*
 338  * Forcibly shutdown the socket causing all listeners to error
 339  */
 340 static void sock_shutdown(struct nbd_device *nbd)
 341 {
 342         struct nbd_config *config = nbd->config;
 343         int i;
 344
 345         if (config->num_connections == 0)
 346                 return;
 347         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 348                 return;
 349
 350         for (i = 0; i < config->num_connections; i++) {
 351                 struct nbd_sock *nsock = config->socks[i];
 352                 mutex_lock(&nsock->tx_lock);
 353                 nbd_mark_nsock_dead(nbd, nsock, 0);
 354                 mutex_unlock(&nsock->tx_lock);
 355         }
 356         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 357 }
 358
 359 static u32 req_to_nbd_cmd_type(struct request *req)
 360 {
 361         switch (req_op(req)) {
 362         case REQ_OP_DISCARD:
 363                 return NBD_CMD_TRIM;
 364         case REQ_OP_FLUSH:
 365                 return NBD_CMD_FLUSH;
 366         case REQ_OP_WRITE:
 367                 return NBD_CMD_WRITE;
 368         case REQ_OP_READ:
 369                 return NBD_CMD_READ;
 370         default:
 371                 return U32_MAX;
 372         }
 373 }
 374
 375 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 376                                                  bool reserved)
 377 {
 378         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 379         struct nbd_device *nbd = cmd->nbd;
 380         struct nbd_config *config;
 381
 382         if (!mutex_trylock(&cmd->lock))
 383                 return BLK_EH_RESET_TIMER;
 384
 385         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 386                 cmd->status = BLK_STS_TIMEOUT;
 387                 mutex_unlock(&cmd->lock);
 388                 goto done;
 389         }
 390         config = nbd->config;
 391
 392         if (config->num_connections > 1 ||
 393             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 394                 dev_err_ratelimited(nbd_to_dev(nbd),
 395                                     "Connection timed out, retrying (%d/%d alive)\n",
 396                                     atomic_read(&config->live_connections),
 397                                     config->num_connections);
 398                 /*
 399                  * Hooray we have more connections, requeue this IO, the submit
 400                  * path will put it on a real connection. Or if only one
 401                  * connection is configured, the submit path will wait util
 402                  * a new connection is reconfigured or util dead timeout.
 403                  */
 404                 if (config->socks) {
 405                         if (cmd->index < config->num_connections) {
 406                                 struct nbd_sock *nsock =
 407                                         config->socks[cmd->index];
 408                                 mutex_lock(&nsock->tx_lock);
 409                                 /* We can have multiple outstanding requests, so
 410                                  * we don't want to mark the nsock dead if we've
 411                                  * already reconnected with a new socket, so
 412                                  * only mark it dead if its the same socket we
 413                                  * were sent out on.
 414                                  */
 415                                 if (cmd->cookie == nsock->cookie)
 416                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 417                                 mutex_unlock(&nsock->tx_lock);
 418                         }
 419                         mutex_unlock(&cmd->lock);
 420                         nbd_requeue_cmd(cmd);
 421                         nbd_config_put(nbd);
 422                         return BLK_EH_DONE;
 423                 }
 424         }
 425
 426         if (!nbd->tag_set.timeout) {
 427                 /*
 428                  * Userspace sets timeout=0 to disable socket disconnection,
 429                  * so just warn and reset the timer.
 430                  */
 431                 struct nbd_sock *nsock = config->socks[cmd->index];
 432                 cmd->retries++;
 433                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 434                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 435                         (unsigned long long)blk_rq_pos(req) << 9,
 436                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 437
 438                 mutex_lock(&nsock->tx_lock);
 439                 if (cmd->cookie != nsock->cookie) {
 440                         nbd_requeue_cmd(cmd);
 441                         mutex_unlock(&nsock->tx_lock);
 442                         mutex_unlock(&cmd->lock);
 443                         nbd_config_put(nbd);
 444                         return BLK_EH_DONE;
 445                 }
 446                 mutex_unlock(&nsock->tx_lock);
 447                 mutex_unlock(&cmd->lock);
 448                 nbd_config_put(nbd);
 449                 return BLK_EH_RESET_TIMER;
 450         }
 451
 452         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 453         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 454         cmd->status = BLK_STS_IOERR;
 455         mutex_unlock(&cmd->lock);
 456         sock_shutdown(nbd);
 457         nbd_config_put(nbd);
 458 done:
 459         blk_mq_complete_request(req);
 460         return BLK_EH_DONE;
 461 }
 462
 463 /*
 464  *  Send or receive packet.
 465  */
 466 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 467                      struct iov_iter *iter, int msg_flags, int *sent)
 468 {
 469         struct nbd_config *config = nbd->config;
 470         struct socket *sock = config->socks[index]->sock;
 471         int result;
 472         struct msghdr msg;
 473         unsigned int noreclaim_flag;
 474
 475         if (unlikely(!sock)) {
 476                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 477                         "Attempted %s on closed socket in sock_xmit\n",
 478                         (send ? "send" : "recv"));
 479                 return -EINVAL;
 480         }
 481
 482         msg.msg_iter = *iter;
 483
 484         noreclaim_flag = memalloc_noreclaim_save();
 485         do {
 486                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 487                 msg.msg_name = NULL;
 488                 msg.msg_namelen = 0;
 489                 msg.msg_control = NULL;
 490                 msg.msg_controllen = 0;
 491                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 492
 493                 if (send)
 494                         result = sock_sendmsg(sock, &msg);
 495                 else
 496                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 497
 498                 if (result <= 0) {
 499                         if (result == 0)
 500                                 result = -EPIPE; /* short read */
 501                         break;
 502                 }
 503                 if (sent)
 504                         *sent += result;
 505         } while (msg_data_left(&msg));
 506
 507         memalloc_noreclaim_restore(noreclaim_flag);
 508
 509         return result;
 510 }
 511
 512 /*
 513  * Different settings for sk->sk_sndtimeo can result in different return values
 514  * if there is a signal pending when we enter sendmsg, because reasons?
 515  */
 516 static inline int was_interrupted(int result)
 517 {
 518         return result == -ERESTARTSYS || result == -EINTR;
 519 }
 520
 521 /* always call with the tx_lock held */
 522 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 523 {
 524         struct request *req = blk_mq_rq_from_pdu(cmd);
 525         struct nbd_config *config = nbd->config;
 526         struct nbd_sock *nsock = config->socks[index];
 527         int result;
 528         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 529         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 530         struct iov_iter from;
 531         unsigned long size = blk_rq_bytes(req);
 532         struct bio *bio;
 533         u64 handle;
 534         u32 type;
 535         u32 nbd_cmd_flags = 0;
 536         int sent = nsock->sent, skip = 0;
 537
 538         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 539
 540         type = req_to_nbd_cmd_type(req);
 541         if (type == U32_MAX)
 542                 return -EIO;
 543
 544         if (rq_data_dir(req) == WRITE &&
 545             (config->flags & NBD_FLAG_READ_ONLY)) {
 546                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 547                                     "Write on read-only\n");
 548                 return -EIO;
 549         }
 550
 551         if (req->cmd_flags & REQ_FUA)
 552                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 553
 554         /* We did a partial send previously, and we at least sent the whole
 555          * request struct, so just go and send the rest of the pages in the
 556          * request.
 557          */
 558         if (sent) {
 559                 if (sent >= sizeof(request)) {
 560                         skip = sent - sizeof(request);
 561
 562                         /* initialize handle for tracing purposes */
 563                         handle = nbd_cmd_handle(cmd);
 564
 565                         goto send_pages;
 566                 }
 567                 iov_iter_advance(&from, sent);
 568         } else {
 569                 cmd->cmd_cookie++;
 570         }
 571         cmd->index = index;
 572         cmd->cookie = nsock->cookie;
 573         cmd->retries = 0;
 574         request.type = htonl(type | nbd_cmd_flags);
 575         if (type != NBD_CMD_FLUSH) {
 576                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 577                 request.len = htonl(size);
 578         }
 579         handle = nbd_cmd_handle(cmd);
 580         memcpy(request.handle, &handle, sizeof(handle));
 581
 582         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 583
 584         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 585                 req, nbdcmd_to_ascii(type),
 586                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 587         result = sock_xmit(nbd, index, 1, &from,
 588                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 589         trace_nbd_header_sent(req, handle);
 590         if (result <= 0) {
 591                 if (was_interrupted(result)) {
 592                         /* If we havne't sent anything we can just return BUSY,
 593                          * however if we have sent something we need to make
 594                          * sure we only allow this req to be sent until we are
 595                          * completely done.
 596                          */
 597                         if (sent) {
 598                                 nsock->pending = req;
 599                                 nsock->sent = sent;
 600                         }
 601                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 602                         return BLK_STS_RESOURCE;
 603                 }
 604                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 605                         "Send control failed (result %d)\n", result);
 606                 return -EAGAIN;
 607         }
 608 send_pages:
 609         if (type != NBD_CMD_WRITE)
 610                 goto out;
 611
 612         bio = req->bio;
 613         while (bio) {
 614                 struct bio *next = bio->bi_next;
 615                 struct bvec_iter iter;
 616                 struct bio_vec bvec;
 617
 618                 bio_for_each_segment(bvec, bio, iter) {
 619                         bool is_last = !next && bio_iter_last(bvec, iter);
 620                         int flags = is_last ? 0 : MSG_MORE;
 621
 622                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 623                                 req, bvec.bv_len);
 624                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 625                         if (skip) {
 626                                 if (skip >= iov_iter_count(&from)) {
 627                                         skip -= iov_iter_count(&from);
 628                                         continue;
 629                                 }
 630                                 iov_iter_advance(&from, skip);
 631                                 skip = 0;
 632                         }
 633                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 634                         if (result <= 0) {
 635                                 if (was_interrupted(result)) {
 636                                         /* We've already sent the header, we
 637                                          * have no choice but to set pending and
 638                                          * return BUSY.
 639                                          */
 640                                         nsock->pending = req;
 641                                         nsock->sent = sent;
 642                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 643                                         return BLK_STS_RESOURCE;
 644                                 }
 645                                 dev_err(disk_to_dev(nbd->disk),
 646                                         "Send data failed (result %d)\n",
 647                                         result);
 648                                 return -EAGAIN;
 649                         }
 650                         /*
 651                          * The completion might already have come in,
 652                          * so break for the last one instead of letting
 653                          * the iterator do it. This prevents use-after-free
 654                          * of the bio.
 655                          */
 656                         if (is_last)
 657                                 break;
 658                 }
 659                 bio = next;
 660         }
 661 out:
 662         trace_nbd_payload_sent(req, handle);
 663         nsock->pending = NULL;
 664         nsock->sent = 0;
 665         return 0;
 666 }
 667
 668 /* NULL returned = something went wrong, inform userspace */
 669 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 670 {
 671         struct nbd_config *config = nbd->config;
 672         int result;
 673         struct nbd_reply reply;
 674         struct nbd_cmd *cmd;
 675         struct request *req = NULL;
 676         u64 handle;
 677         u16 hwq;
 678         u32 tag;
 679         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 680         struct iov_iter to;
 681         int ret = 0;
 682
 683         reply.magic = 0;
 684         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 685         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 686         if (result <= 0) {
 687                 if (!nbd_disconnected(config))
 688                         dev_err(disk_to_dev(nbd->disk),
 689                                 "Receive control failed (result %d)\n", result);
 690                 return ERR_PTR(result);
 691         }
 692
 693         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 694                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 695                                 (unsigned long)ntohl(reply.magic));
 696                 return ERR_PTR(-EPROTO);
 697         }
 698
 699         memcpy(&handle, reply.handle, sizeof(handle));
 700         tag = nbd_handle_to_tag(handle);
 701         hwq = blk_mq_unique_tag_to_hwq(tag);
 702         if (hwq < nbd->tag_set.nr_hw_queues)
 703                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 704                                        blk_mq_unique_tag_to_tag(tag));
 705         if (!req || !blk_mq_request_started(req)) {
 706                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 707                         tag, req);
 708                 return ERR_PTR(-ENOENT);
 709         }
 710         trace_nbd_header_received(req, handle);
 711         cmd = blk_mq_rq_to_pdu(req);
 712
 713         mutex_lock(&cmd->lock);
 714         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 715                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 716                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 717                 ret = -ENOENT;
 718                 goto out;
 719         }
 720         if (cmd->status != BLK_STS_OK) {
 721                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 722                         req);
 723                 ret = -ENOENT;
 724                 goto out;
 725         }
 726         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 727                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 728                         req);
 729                 ret = -ENOENT;
 730                 goto out;
 731         }
 732         if (ntohl(reply.error)) {
 733                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 734                         ntohl(reply.error));
 735                 cmd->status = BLK_STS_IOERR;
 736                 goto out;
 737         }
 738
 739         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 740         if (rq_data_dir(req) != WRITE) {
 741                 struct req_iterator iter;
 742                 struct bio_vec bvec;
 743
 744                 rq_for_each_segment(bvec, req, iter) {
 745                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 746                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 747                         if (result <= 0) {
 748                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 749                                         result);
 750                                 /*
 751                                  * If we've disconnected, we need to make sure we
 752                                  * complete this request, otherwise error out
 753                                  * and let the timeout stuff handle resubmitting
 754                                  * this request onto another connection.
 755                                  */
 756                                 if (nbd_disconnected(config)) {
 757                                         cmd->status = BLK_STS_IOERR;
 758                                         goto out;
 759                                 }
 760                                 ret = -EIO;
 761                                 goto out;
 762                         }
 763                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 764                                 req, bvec.bv_len);
 765                 }
 766         }
 767 out:
 768         trace_nbd_payload_received(req, handle);
 769         mutex_unlock(&cmd->lock);
 770         return ret ? ERR_PTR(ret) : cmd;
 771 }
 772
 773 static void recv_work(struct work_struct *work)
 774 {
 775         struct recv_thread_args *args = container_of(work,
 776                                                      struct recv_thread_args,
 777                                                      work);
 778         struct nbd_device *nbd = args->nbd;
 779         struct nbd_config *config = nbd->config;
 780         struct nbd_cmd *cmd;
 781         struct request *rq;
 782
 783         while (1) {
 784                 cmd = nbd_read_stat(nbd, args->index);
 785                 if (IS_ERR(cmd)) {
 786                         struct nbd_sock *nsock = config->socks[args->index];
 787
 788                         mutex_lock(&nsock->tx_lock);
 789                         nbd_mark_nsock_dead(nbd, nsock, 1);
 790                         mutex_unlock(&nsock->tx_lock);
 791                         break;
 792                 }
 793
 794                 rq = blk_mq_rq_from_pdu(cmd);
 795                 if (likely(!blk_should_fake_timeout(rq->q)))
 796                         blk_mq_complete_request(rq);
 797         }
 798         nbd_config_put(nbd);
 799         atomic_dec(&config->recv_threads);
 800         wake_up(&config->recv_wq);
 801         kfree(args);
 802 }
 803
 804 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 805 {
 806         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 807
 808         mutex_lock(&cmd->lock);
 809         cmd->status = BLK_STS_IOERR;
 810         mutex_unlock(&cmd->lock);
 811
 812         blk_mq_complete_request(req);
 813         return true;
 814 }
 815
 816 static void nbd_clear_que(struct nbd_device *nbd)
 817 {
 818         blk_mq_quiesce_queue(nbd->disk->queue);
 819         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 820         blk_mq_unquiesce_queue(nbd->disk->queue);
 821         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 822 }
 823
 824 static int find_fallback(struct nbd_device *nbd, int index)
 825 {
 826         struct nbd_config *config = nbd->config;
 827         int new_index = -1;
 828         struct nbd_sock *nsock = config->socks[index];
 829         int fallback = nsock->fallback_index;
 830
 831         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 832                 return new_index;
 833
 834         if (config->num_connections <= 1) {
 835                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 836                                     "Dead connection, failed to find a fallback\n");
 837                 return new_index;
 838         }
 839
 840         if (fallback >= 0 && fallback < config->num_connections &&
 841             !config->socks[fallback]->dead)
 842                 return fallback;
 843
 844         if (nsock->fallback_index < 0 ||
 845             nsock->fallback_index >= config->num_connections ||
 846             config->socks[nsock->fallback_index]->dead) {
 847                 int i;
 848                 for (i = 0; i < config->num_connections; i++) {
 849                         if (i == index)
 850                                 continue;
 851                         if (!config->socks[i]->dead) {
 852                                 new_index = i;
 853                                 break;
 854                         }
 855                 }
 856                 nsock->fallback_index = new_index;
 857                 if (new_index < 0) {
 858                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 859                                             "Dead connection, failed to find a fallback\n");
 860                         return new_index;
 861                 }
 862         }
 863         new_index = nsock->fallback_index;
 864         return new_index;
 865 }
 866
 867 static int wait_for_reconnect(struct nbd_device *nbd)
 868 {
 869         struct nbd_config *config = nbd->config;
 870         if (!config->dead_conn_timeout)
 871                 return 0;
 872         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 873                 return 0;
 874         return wait_event_timeout(config->conn_wait,
 875                                   atomic_read(&config->live_connections) > 0,
 876                                   config->dead_conn_timeout) > 0;
 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         set_capacity(bdev->bd_disk, 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         return nbd_set_size(nbd, config->bytesize, config->blksize);
1313 }
1314
1315 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1316 {
1317         struct nbd_config *config = nbd->config;
1318         int ret;
1319
1320         ret = nbd_start_device(nbd);
1321         if (ret)
1322                 return ret;
1323
1324         if (max_part)
1325                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1326         mutex_unlock(&nbd->config_lock);
1327         ret = wait_event_interruptible(config->recv_wq,
1328                                          atomic_read(&config->recv_threads) == 0);
1329         if (ret)
1330                 sock_shutdown(nbd);
1331         flush_workqueue(nbd->recv_workq);
1332
1333         mutex_lock(&nbd->config_lock);
1334         nbd_bdev_reset(bdev);
1335         /* user requested, ignore socket errors */
1336         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1337                 ret = 0;
1338         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1339                 ret = -ETIMEDOUT;
1340         return ret;
1341 }
1342
1343 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1344                                  struct block_device *bdev)
1345 {
1346         sock_shutdown(nbd);
1347         __invalidate_device(bdev, true);
1348         nbd_bdev_reset(bdev);
1349         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1350                                &nbd->config->runtime_flags))
1351                 nbd_config_put(nbd);
1352 }
1353
1354 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1355 {
1356         nbd->tag_set.timeout = timeout * HZ;
1357         if (timeout)
1358                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1359         else
1360                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1361 }
1362
1363 /* Must be called with config_lock held */
1364 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1365                        unsigned int cmd, unsigned long arg)
1366 {
1367         struct nbd_config *config = nbd->config;
1368
1369         switch (cmd) {
1370         case NBD_DISCONNECT:
1371                 return nbd_disconnect(nbd);
1372         case NBD_CLEAR_SOCK:
1373                 nbd_clear_sock_ioctl(nbd, bdev);
1374                 return 0;
1375         case NBD_SET_SOCK:
1376                 return nbd_add_socket(nbd, arg, false);
1377         case NBD_SET_BLKSIZE:
1378                 return nbd_set_size(nbd, config->bytesize, arg);
1379         case NBD_SET_SIZE:
1380                 return nbd_set_size(nbd, arg, config->blksize);
1381         case NBD_SET_SIZE_BLOCKS:
1382                 return nbd_set_size(nbd, arg * config->blksize,
1383                                     config->blksize);
1384         case NBD_SET_TIMEOUT:
1385                 nbd_set_cmd_timeout(nbd, arg);
1386                 return 0;
1387
1388         case NBD_SET_FLAGS:
1389                 config->flags = arg;
1390                 return 0;
1391         case NBD_DO_IT:
1392                 return nbd_start_device_ioctl(nbd, bdev);
1393         case NBD_CLEAR_QUE:
1394                 /*
1395                  * This is for compatibility only.  The queue is always cleared
1396                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1397                  */
1398                 return 0;
1399         case NBD_PRINT_DEBUG:
1400                 /*
1401                  * For compatibility only, we no longer keep a list of
1402                  * outstanding requests.
1403                  */
1404                 return 0;
1405         }
1406         return -ENOTTY;
1407 }
1408
1409 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1410                      unsigned int cmd, unsigned long arg)
1411 {
1412         struct nbd_device *nbd = bdev->bd_disk->private_data;
1413         struct nbd_config *config = nbd->config;
1414         int error = -EINVAL;
1415
1416         if (!capable(CAP_SYS_ADMIN))
1417                 return -EPERM;
1418
1419         /* The block layer will pass back some non-nbd ioctls in case we have
1420          * special handling for them, but we don't so just return an error.
1421          */
1422         if (_IOC_TYPE(cmd) != 0xab)
1423                 return -EINVAL;
1424
1425         mutex_lock(&nbd->config_lock);
1426
1427         /* Don't allow ioctl operations on a nbd device that was created with
1428          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1429          */
1430         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1431             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1432                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1433         else
1434                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1435         mutex_unlock(&nbd->config_lock);
1436         return error;
1437 }
1438
1439 static struct nbd_config *nbd_alloc_config(void)
1440 {
1441         struct nbd_config *config;
1442
1443         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1444         if (!config)
1445                 return NULL;
1446         atomic_set(&config->recv_threads, 0);
1447         init_waitqueue_head(&config->recv_wq);
1448         init_waitqueue_head(&config->conn_wait);
1449         config->blksize = NBD_DEF_BLKSIZE;
1450         atomic_set(&config->live_connections, 0);
1451         try_module_get(THIS_MODULE);
1452         return config;
1453 }
1454
1455 static int nbd_open(struct block_device *bdev, fmode_t mode)
1456 {
1457         struct nbd_device *nbd;
1458         int ret = 0;
1459
1460         mutex_lock(&nbd_index_mutex);
1461         nbd = bdev->bd_disk->private_data;
1462         if (!nbd) {
1463                 ret = -ENXIO;
1464                 goto out;
1465         }
1466         if (!refcount_inc_not_zero(&nbd->refs)) {
1467                 ret = -ENXIO;
1468                 goto out;
1469         }
1470         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1471                 struct nbd_config *config;
1472
1473                 mutex_lock(&nbd->config_lock);
1474                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1475                         mutex_unlock(&nbd->config_lock);
1476                         goto out;
1477                 }
1478                 config = nbd->config = nbd_alloc_config();
1479                 if (!config) {
1480                         ret = -ENOMEM;
1481                         mutex_unlock(&nbd->config_lock);
1482                         goto out;
1483                 }
1484                 refcount_set(&nbd->config_refs, 1);
1485                 refcount_inc(&nbd->refs);
1486                 mutex_unlock(&nbd->config_lock);
1487                 if (max_part)
1488                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1489         } else if (nbd_disconnected(nbd->config)) {
1490                 if (max_part)
1491                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1492         }
1493 out:
1494         mutex_unlock(&nbd_index_mutex);
1495         return ret;
1496 }
1497
1498 static void nbd_release(struct gendisk *disk, fmode_t mode)
1499 {
1500         struct nbd_device *nbd = disk->private_data;
1501
1502         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1503                         disk->part0->bd_openers == 0)
1504                 nbd_disconnect_and_put(nbd);
1505
1506         nbd_config_put(nbd);
1507         nbd_put(nbd);
1508 }
1509
1510 static const struct block_device_operations nbd_fops =
1511 {
1512         .owner =        THIS_MODULE,
1513         .open =         nbd_open,
1514         .release =      nbd_release,
1515         .ioctl =        nbd_ioctl,
1516         .compat_ioctl = nbd_ioctl,
1517 };
1518
1519 #if IS_ENABLED(CONFIG_DEBUG_FS)
1520
1521 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1522 {
1523         struct nbd_device *nbd = s->private;
1524
1525         if (nbd->task_recv)
1526                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1527
1528         return 0;
1529 }
1530
1531 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1532
1533 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1534 {
1535         struct nbd_device *nbd = s->private;
1536         u32 flags = nbd->config->flags;
1537
1538         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1539
1540         seq_puts(s, "Known flags:\n");
1541
1542         if (flags & NBD_FLAG_HAS_FLAGS)
1543                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1544         if (flags & NBD_FLAG_READ_ONLY)
1545                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1546         if (flags & NBD_FLAG_SEND_FLUSH)
1547                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1548         if (flags & NBD_FLAG_SEND_FUA)
1549                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1550         if (flags & NBD_FLAG_SEND_TRIM)
1551                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1552
1553         return 0;
1554 }
1555
1556 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1557
1558 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1559 {
1560         struct dentry *dir;
1561         struct nbd_config *config = nbd->config;
1562
1563         if (!nbd_dbg_dir)
1564                 return -EIO;
1565
1566         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1567         if (!dir) {
1568                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1569                         nbd_name(nbd));
1570                 return -EIO;
1571         }
1572         config->dbg_dir = dir;
1573
1574         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1575         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1576         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1577         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1578         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1579
1580         return 0;
1581 }
1582
1583 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1584 {
1585         debugfs_remove_recursive(nbd->config->dbg_dir);
1586 }
1587
1588 static int nbd_dbg_init(void)
1589 {
1590         struct dentry *dbg_dir;
1591
1592         dbg_dir = debugfs_create_dir("nbd", NULL);
1593         if (!dbg_dir)
1594                 return -EIO;
1595
1596         nbd_dbg_dir = dbg_dir;
1597
1598         return 0;
1599 }
1600
1601 static void nbd_dbg_close(void)
1602 {
1603         debugfs_remove_recursive(nbd_dbg_dir);
1604 }
1605
1606 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1607
1608 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1609 {
1610         return 0;
1611 }
1612
1613 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1614 {
1615 }
1616
1617 static int nbd_dbg_init(void)
1618 {
1619         return 0;
1620 }
1621
1622 static void nbd_dbg_close(void)
1623 {
1624 }
1625
1626 #endif
1627
1628 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1629                             unsigned int hctx_idx, unsigned int numa_node)
1630 {
1631         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1632         cmd->nbd = set->driver_data;
1633         cmd->flags = 0;
1634         mutex_init(&cmd->lock);
1635         return 0;
1636 }
1637
1638 static const struct blk_mq_ops nbd_mq_ops = {
1639         .queue_rq       = nbd_queue_rq,
1640         .complete       = nbd_complete_rq,
1641         .init_request   = nbd_init_request,
1642         .timeout        = nbd_xmit_timeout,
1643 };
1644
1645 static int nbd_dev_add(int index)
1646 {
1647         struct nbd_device *nbd;
1648         struct gendisk *disk;
1649         struct request_queue *q;
1650         int err = -ENOMEM;
1651
1652         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1653         if (!nbd)
1654                 goto out;
1655
1656         disk = alloc_disk(1 << part_shift);
1657         if (!disk)
1658                 goto out_free_nbd;
1659
1660         if (index >= 0) {
1661                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1662                                 GFP_KERNEL);
1663                 if (err == -ENOSPC)
1664                         err = -EEXIST;
1665         } else {
1666                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1667                 if (err >= 0)
1668                         index = err;
1669         }
1670         if (err < 0)
1671                 goto out_free_disk;
1672
1673         nbd->index = index;
1674         nbd->disk = disk;
1675         nbd->tag_set.ops = &nbd_mq_ops;
1676         nbd->tag_set.nr_hw_queues = 1;
1677         nbd->tag_set.queue_depth = 128;
1678         nbd->tag_set.numa_node = NUMA_NO_NODE;
1679         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1680         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1681                 BLK_MQ_F_BLOCKING;
1682         nbd->tag_set.driver_data = nbd;
1683         nbd->destroy_complete = NULL;
1684
1685         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1686         if (err)
1687                 goto out_free_idr;
1688
1689         q = blk_mq_init_queue(&nbd->tag_set);
1690         if (IS_ERR(q)) {
1691                 err = PTR_ERR(q);
1692                 goto out_free_tags;
1693         }
1694         disk->queue = q;
1695
1696         /*
1697          * Tell the block layer that we are not a rotational device
1698          */
1699         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1700         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1701         disk->queue->limits.discard_granularity = 0;
1702         disk->queue->limits.discard_alignment = 0;
1703         blk_queue_max_discard_sectors(disk->queue, 0);
1704         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1705         blk_queue_max_segments(disk->queue, USHRT_MAX);
1706         blk_queue_max_hw_sectors(disk->queue, 65536);
1707         disk->queue->limits.max_sectors = 256;
1708
1709         mutex_init(&nbd->config_lock);
1710         refcount_set(&nbd->config_refs, 0);
1711         refcount_set(&nbd->refs, 1);
1712         INIT_LIST_HEAD(&nbd->list);
1713         disk->major = NBD_MAJOR;
1714         disk->first_minor = index << part_shift;
1715         disk->fops = &nbd_fops;
1716         disk->private_data = nbd;
1717         sprintf(disk->disk_name, "nbd%d", index);
1718         add_disk(disk);
1719         nbd_total_devices++;
1720         return index;
1721
1722 out_free_tags:
1723         blk_mq_free_tag_set(&nbd->tag_set);
1724 out_free_idr:
1725         idr_remove(&nbd_index_idr, index);
1726 out_free_disk:
1727         put_disk(disk);
1728 out_free_nbd:
1729         kfree(nbd);
1730 out:
1731         return err;
1732 }
1733
1734 static int find_free_cb(int id, void *ptr, void *data)
1735 {
1736         struct nbd_device *nbd = ptr;
1737         struct nbd_device **found = data;
1738
1739         if (!refcount_read(&nbd->config_refs)) {
1740                 *found = nbd;
1741                 return 1;
1742         }
1743         return 0;
1744 }
1745
1746 /* Netlink interface. */
1747 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1748         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1749         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1750         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1751         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1752         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1753         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1754         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1755         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1756         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1757 };
1758
1759 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1760         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1761 };
1762
1763 /* We don't use this right now since we don't parse the incoming list, but we
1764  * still want it here so userspace knows what to expect.
1765  */
1766 static const struct nla_policy __attribute__((unused))
1767 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1768         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1769         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1770 };
1771
1772 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1773 {
1774         struct nbd_config *config = nbd->config;
1775         u64 bsize = config->blksize;
1776         u64 bytes = config->bytesize;
1777
1778         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1779                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1780
1781         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1782                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1783
1784         if (bytes != config->bytesize || bsize != config->blksize)
1785                 return nbd_set_size(nbd, bytes, bsize);
1786         return 0;
1787 }
1788
1789 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1790 {
1791         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1792         struct nbd_device *nbd = NULL;
1793         struct nbd_config *config;
1794         int index = -1;
1795         int ret;
1796         bool put_dev = false;
1797
1798         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1799                 return -EPERM;
1800
1801         if (info->attrs[NBD_ATTR_INDEX])
1802                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1803         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1804                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1805                 return -EINVAL;
1806         }
1807         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1808                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1809                 return -EINVAL;
1810         }
1811 again:
1812         mutex_lock(&nbd_index_mutex);
1813         if (index == -1) {
1814                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1815                 if (ret == 0) {
1816                         int new_index;
1817                         new_index = nbd_dev_add(-1);
1818                         if (new_index < 0) {
1819                                 mutex_unlock(&nbd_index_mutex);
1820                                 printk(KERN_ERR "nbd: failed to add new device\n");
1821                                 return new_index;
1822                         }
1823                         nbd = idr_find(&nbd_index_idr, new_index);
1824                 }
1825         } else {
1826                 nbd = idr_find(&nbd_index_idr, index);
1827                 if (!nbd) {
1828                         ret = nbd_dev_add(index);
1829                         if (ret < 0) {
1830                                 mutex_unlock(&nbd_index_mutex);
1831                                 printk(KERN_ERR "nbd: failed to add new device\n");
1832                                 return ret;
1833                         }
1834                         nbd = idr_find(&nbd_index_idr, index);
1835                 }
1836         }
1837         if (!nbd) {
1838                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1839                        index);
1840                 mutex_unlock(&nbd_index_mutex);
1841                 return -EINVAL;
1842         }
1843
1844         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1845             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1846                 nbd->destroy_complete = &destroy_complete;
1847                 mutex_unlock(&nbd_index_mutex);
1848
1849                 /* Wait untill the the nbd stuff is totally destroyed */
1850                 wait_for_completion(&destroy_complete);
1851                 goto again;
1852         }
1853
1854         if (!refcount_inc_not_zero(&nbd->refs)) {
1855                 mutex_unlock(&nbd_index_mutex);
1856                 if (index == -1)
1857                         goto again;
1858                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1859                        index);
1860                 return -EINVAL;
1861         }
1862         mutex_unlock(&nbd_index_mutex);
1863
1864         mutex_lock(&nbd->config_lock);
1865         if (refcount_read(&nbd->config_refs)) {
1866                 mutex_unlock(&nbd->config_lock);
1867                 nbd_put(nbd);
1868                 if (index == -1)
1869                         goto again;
1870                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1871                 return -EBUSY;
1872         }
1873         if (WARN_ON(nbd->config)) {
1874                 mutex_unlock(&nbd->config_lock);
1875                 nbd_put(nbd);
1876                 return -EINVAL;
1877         }
1878         config = nbd->config = nbd_alloc_config();
1879         if (!nbd->config) {
1880                 mutex_unlock(&nbd->config_lock);
1881                 nbd_put(nbd);
1882                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1883                 return -ENOMEM;
1884         }
1885         refcount_set(&nbd->config_refs, 1);
1886         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1887
1888         ret = nbd_genl_size_set(info, nbd);
1889         if (ret)
1890                 goto out;
1891
1892         if (info->attrs[NBD_ATTR_TIMEOUT])
1893                 nbd_set_cmd_timeout(nbd,
1894                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1895         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1896                 config->dead_conn_timeout =
1897                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1898                 config->dead_conn_timeout *= HZ;
1899         }
1900         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1901                 config->flags =
1902                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1903         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1904                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1905                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1906                         /*
1907                          * We have 1 ref to keep the device around, and then 1
1908                          * ref for our current operation here, which will be
1909                          * inherited by the config.  If we already have
1910                          * DESTROY_ON_DISCONNECT set then we know we don't have
1911                          * that extra ref already held so we don't need the
1912                          * put_dev.
1913                          */
1914                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1915                                               &nbd->flags))
1916                                 put_dev = true;
1917                 } else {
1918                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1919                                                &nbd->flags))
1920                                 refcount_inc(&nbd->refs);
1921                 }
1922                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1923                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1924                                 &config->runtime_flags);
1925                 }
1926         }
1927
1928         if (info->attrs[NBD_ATTR_SOCKETS]) {
1929                 struct nlattr *attr;
1930                 int rem, fd;
1931
1932                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1933                                     rem) {
1934                         struct nlattr *socks[NBD_SOCK_MAX+1];
1935
1936                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1937                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1938                                 ret = -EINVAL;
1939                                 goto out;
1940                         }
1941                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1942                                                           attr,
1943                                                           nbd_sock_policy,
1944                                                           info->extack);
1945                         if (ret != 0) {
1946                                 printk(KERN_ERR "nbd: error processing sock list\n");
1947                                 ret = -EINVAL;
1948                                 goto out;
1949                         }
1950                         if (!socks[NBD_SOCK_FD])
1951                                 continue;
1952                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1953                         ret = nbd_add_socket(nbd, fd, true);
1954                         if (ret)
1955                                 goto out;
1956                 }
1957         }
1958         ret = nbd_start_device(nbd);
1959 out:
1960         mutex_unlock(&nbd->config_lock);
1961         if (!ret) {
1962                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
1963                 refcount_inc(&nbd->config_refs);
1964                 nbd_connect_reply(info, nbd->index);
1965         }
1966         nbd_config_put(nbd);
1967         if (put_dev)
1968                 nbd_put(nbd);
1969         return ret;
1970 }
1971
1972 static void nbd_disconnect_and_put(struct nbd_device *nbd)
1973 {
1974         mutex_lock(&nbd->config_lock);
1975         nbd_disconnect(nbd);
1976         nbd_clear_sock(nbd);
1977         mutex_unlock(&nbd->config_lock);
1978         /*
1979          * Make sure recv thread has finished, so it does not drop the last
1980          * config ref and try to destroy the workqueue from inside the work
1981          * queue.
1982          */
1983         if (nbd->recv_workq)
1984                 flush_workqueue(nbd->recv_workq);
1985         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1986                                &nbd->config->runtime_flags))
1987                 nbd_config_put(nbd);
1988 }
1989
1990 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
1991 {
1992         struct nbd_device *nbd;
1993         int index;
1994
1995         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1996                 return -EPERM;
1997
1998         if (!info->attrs[NBD_ATTR_INDEX]) {
1999                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2000                 return -EINVAL;
2001         }
2002         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2003         mutex_lock(&nbd_index_mutex);
2004         nbd = idr_find(&nbd_index_idr, index);
2005         if (!nbd) {
2006                 mutex_unlock(&nbd_index_mutex);
2007                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2008                        index);
2009                 return -EINVAL;
2010         }
2011         if (!refcount_inc_not_zero(&nbd->refs)) {
2012                 mutex_unlock(&nbd_index_mutex);
2013                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2014                        index);
2015                 return -EINVAL;
2016         }
2017         mutex_unlock(&nbd_index_mutex);
2018         if (!refcount_inc_not_zero(&nbd->config_refs))
2019                 goto put_nbd;
2020         nbd_disconnect_and_put(nbd);
2021         nbd_config_put(nbd);
2022 put_nbd:
2023         nbd_put(nbd);
2024         return 0;
2025 }
2026
2027 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2028 {
2029         struct nbd_device *nbd = NULL;
2030         struct nbd_config *config;
2031         int index;
2032         int ret = 0;
2033         bool put_dev = false;
2034
2035         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2036                 return -EPERM;
2037
2038         if (!info->attrs[NBD_ATTR_INDEX]) {
2039                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2040                 return -EINVAL;
2041         }
2042         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2043         mutex_lock(&nbd_index_mutex);
2044         nbd = idr_find(&nbd_index_idr, index);
2045         if (!nbd) {
2046                 mutex_unlock(&nbd_index_mutex);
2047                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2048                        index);
2049                 return -EINVAL;
2050         }
2051         if (!refcount_inc_not_zero(&nbd->refs)) {
2052                 mutex_unlock(&nbd_index_mutex);
2053                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2054                        index);
2055                 return -EINVAL;
2056         }
2057         mutex_unlock(&nbd_index_mutex);
2058
2059         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2060                 dev_err(nbd_to_dev(nbd),
2061                         "not configured, cannot reconfigure\n");
2062                 nbd_put(nbd);
2063                 return -EINVAL;
2064         }
2065
2066         mutex_lock(&nbd->config_lock);
2067         config = nbd->config;
2068         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2069             !nbd->task_recv) {
2070                 dev_err(nbd_to_dev(nbd),
2071                         "not configured, cannot reconfigure\n");
2072                 ret = -EINVAL;
2073                 goto out;
2074         }
2075
2076         ret = nbd_genl_size_set(info, nbd);
2077         if (ret)
2078                 goto out;
2079
2080         if (info->attrs[NBD_ATTR_TIMEOUT])
2081                 nbd_set_cmd_timeout(nbd,
2082                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2083         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2084                 config->dead_conn_timeout =
2085                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2086                 config->dead_conn_timeout *= HZ;
2087         }
2088         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2089                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2090                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2091                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2092                                               &nbd->flags))
2093                                 put_dev = true;
2094                 } else {
2095                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2096                                                &nbd->flags))
2097                                 refcount_inc(&nbd->refs);
2098                 }
2099
2100                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2101                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2102                                         &config->runtime_flags);
2103                 } else {
2104                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2105                                         &config->runtime_flags);
2106                 }
2107         }
2108
2109         if (info->attrs[NBD_ATTR_SOCKETS]) {
2110                 struct nlattr *attr;
2111                 int rem, fd;
2112
2113                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2114                                     rem) {
2115                         struct nlattr *socks[NBD_SOCK_MAX+1];
2116
2117                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2118                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2119                                 ret = -EINVAL;
2120                                 goto out;
2121                         }
2122                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2123                                                           attr,
2124                                                           nbd_sock_policy,
2125                                                           info->extack);
2126                         if (ret != 0) {
2127                                 printk(KERN_ERR "nbd: error processing sock list\n");
2128                                 ret = -EINVAL;
2129                                 goto out;
2130                         }
2131                         if (!socks[NBD_SOCK_FD])
2132                                 continue;
2133                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2134                         ret = nbd_reconnect_socket(nbd, fd);
2135                         if (ret) {
2136                                 if (ret == -ENOSPC)
2137                                         ret = 0;
2138                                 goto out;
2139                         }
2140                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2141                 }
2142         }
2143 out:
2144         mutex_unlock(&nbd->config_lock);
2145         nbd_config_put(nbd);
2146         nbd_put(nbd);
2147         if (put_dev)
2148                 nbd_put(nbd);
2149         return ret;
2150 }
2151
2152 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2153         {
2154                 .cmd    = NBD_CMD_CONNECT,
2155                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2156                 .doit   = nbd_genl_connect,
2157         },
2158         {
2159                 .cmd    = NBD_CMD_DISCONNECT,
2160                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2161                 .doit   = nbd_genl_disconnect,
2162         },
2163         {
2164                 .cmd    = NBD_CMD_RECONFIGURE,
2165                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2166                 .doit   = nbd_genl_reconfigure,
2167         },
2168         {
2169                 .cmd    = NBD_CMD_STATUS,
2170                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2171                 .doit   = nbd_genl_status,
2172         },
2173 };
2174
2175 static const struct genl_multicast_group nbd_mcast_grps[] = {
2176         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2177 };
2178
2179 static struct genl_family nbd_genl_family __ro_after_init = {
2180         .hdrsize        = 0,
2181         .name           = NBD_GENL_FAMILY_NAME,
2182         .version        = NBD_GENL_VERSION,
2183         .module         = THIS_MODULE,
2184         .small_ops      = nbd_connect_genl_ops,
2185         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2186         .maxattr        = NBD_ATTR_MAX,
2187         .policy = nbd_attr_policy,
2188         .mcgrps         = nbd_mcast_grps,
2189         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2190 };
2191
2192 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2193 {
2194         struct nlattr *dev_opt;
2195         u8 connected = 0;
2196         int ret;
2197
2198         /* This is a little racey, but for status it's ok.  The
2199          * reason we don't take a ref here is because we can't
2200          * take a ref in the index == -1 case as we would need
2201          * to put under the nbd_index_mutex, which could
2202          * deadlock if we are configured to remove ourselves
2203          * once we're disconnected.
2204          */
2205         if (refcount_read(&nbd->config_refs))
2206                 connected = 1;
2207         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2208         if (!dev_opt)
2209                 return -EMSGSIZE;
2210         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2211         if (ret)
2212                 return -EMSGSIZE;
2213         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2214                          connected);
2215         if (ret)
2216                 return -EMSGSIZE;
2217         nla_nest_end(reply, dev_opt);
2218         return 0;
2219 }
2220
2221 static int status_cb(int id, void *ptr, void *data)
2222 {
2223         struct nbd_device *nbd = ptr;
2224         return populate_nbd_status(nbd, (struct sk_buff *)data);
2225 }
2226
2227 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2228 {
2229         struct nlattr *dev_list;
2230         struct sk_buff *reply;
2231         void *reply_head;
2232         size_t msg_size;
2233         int index = -1;
2234         int ret = -ENOMEM;
2235
2236         if (info->attrs[NBD_ATTR_INDEX])
2237                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2238
2239         mutex_lock(&nbd_index_mutex);
2240
2241         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2242                                   nla_attr_size(sizeof(u8)));
2243         msg_size *= (index == -1) ? nbd_total_devices : 1;
2244
2245         reply = genlmsg_new(msg_size, GFP_KERNEL);
2246         if (!reply)
2247                 goto out;
2248         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2249                                        NBD_CMD_STATUS);
2250         if (!reply_head) {
2251                 nlmsg_free(reply);
2252                 goto out;
2253         }
2254
2255         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2256         if (index == -1) {
2257                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2258                 if (ret) {
2259                         nlmsg_free(reply);
2260                         goto out;
2261                 }
2262         } else {
2263                 struct nbd_device *nbd;
2264                 nbd = idr_find(&nbd_index_idr, index);
2265                 if (nbd) {
2266                         ret = populate_nbd_status(nbd, reply);
2267                         if (ret) {
2268                                 nlmsg_free(reply);
2269                                 goto out;
2270                         }
2271                 }
2272         }
2273         nla_nest_end(reply, dev_list);
2274         genlmsg_end(reply, reply_head);
2275         ret = genlmsg_reply(reply, info);
2276 out:
2277         mutex_unlock(&nbd_index_mutex);
2278         return ret;
2279 }
2280
2281 static void nbd_connect_reply(struct genl_info *info, int index)
2282 {
2283         struct sk_buff *skb;
2284         void *msg_head;
2285         int ret;
2286
2287         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2288         if (!skb)
2289                 return;
2290         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2291                                      NBD_CMD_CONNECT);
2292         if (!msg_head) {
2293                 nlmsg_free(skb);
2294                 return;
2295         }
2296         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2297         if (ret) {
2298                 nlmsg_free(skb);
2299                 return;
2300         }
2301         genlmsg_end(skb, msg_head);
2302         genlmsg_reply(skb, info);
2303 }
2304
2305 static void nbd_mcast_index(int index)
2306 {
2307         struct sk_buff *skb;
2308         void *msg_head;
2309         int ret;
2310
2311         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2312         if (!skb)
2313                 return;
2314         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2315                                      NBD_CMD_LINK_DEAD);
2316         if (!msg_head) {
2317                 nlmsg_free(skb);
2318                 return;
2319         }
2320         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2321         if (ret) {
2322                 nlmsg_free(skb);
2323                 return;
2324         }
2325         genlmsg_end(skb, msg_head);
2326         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2327 }
2328
2329 static void nbd_dead_link_work(struct work_struct *work)
2330 {
2331         struct link_dead_args *args = container_of(work, struct link_dead_args,
2332                                                    work);
2333         nbd_mcast_index(args->index);
2334         kfree(args);
2335 }
2336
2337 static int __init nbd_init(void)
2338 {
2339         int i;
2340
2341         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2342
2343         if (max_part < 0) {
2344                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2345                 return -EINVAL;
2346         }
2347
2348         part_shift = 0;
2349         if (max_part > 0) {
2350                 part_shift = fls(max_part);
2351
2352                 /*
2353                  * Adjust max_part according to part_shift as it is exported
2354                  * to user space so that user can know the max number of
2355                  * partition kernel should be able to manage.
2356                  *
2357                  * Note that -1 is required because partition 0 is reserved
2358                  * for the whole disk.
2359                  */
2360                 max_part = (1UL << part_shift) - 1;
2361         }
2362
2363         if ((1UL << part_shift) > DISK_MAX_PARTS)
2364                 return -EINVAL;
2365
2366         if (nbds_max > 1UL << (MINORBITS - part_shift))
2367                 return -EINVAL;
2368
2369         if (register_blkdev(NBD_MAJOR, "nbd"))
2370                 return -EIO;
2371
2372         if (genl_register_family(&nbd_genl_family)) {
2373                 unregister_blkdev(NBD_MAJOR, "nbd");
2374                 return -EINVAL;
2375         }
2376         nbd_dbg_init();
2377
2378         mutex_lock(&nbd_index_mutex);
2379         for (i = 0; i < nbds_max; i++)
2380                 nbd_dev_add(i);
2381         mutex_unlock(&nbd_index_mutex);
2382         return 0;
2383 }
2384
2385 static int nbd_exit_cb(int id, void *ptr, void *data)
2386 {
2387         struct list_head *list = (struct list_head *)data;
2388         struct nbd_device *nbd = ptr;
2389
2390         list_add_tail(&nbd->list, list);
2391         return 0;
2392 }
2393
2394 static void __exit nbd_cleanup(void)
2395 {
2396         struct nbd_device *nbd;
2397         LIST_HEAD(del_list);
2398
2399         nbd_dbg_close();
2400
2401         mutex_lock(&nbd_index_mutex);
2402         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2403         mutex_unlock(&nbd_index_mutex);
2404
2405         while (!list_empty(&del_list)) {
2406                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2407                 list_del_init(&nbd->list);
2408                 if (refcount_read(&nbd->refs) != 1)
2409                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2410                 nbd_put(nbd);
2411         }
2412
2413         idr_destroy(&nbd_index_idr);
2414         genl_unregister_family(&nbd_genl_family);
2415         unregister_blkdev(NBD_MAJOR, "nbd");
2416 }
2417
2418 module_init(nbd_init);
2419 module_exit(nbd_cleanup);
2420
2421 MODULE_DESCRIPTION("Network Block Device");
2422 MODULE_LICENSE("GPL");
2423
2424 module_param(nbds_max, int, 0444);
2425 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2426 module_param(max_part, int, 0444);
2427 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");