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
 223         if (disk) {
 224                 del_gendisk(disk);
 225                 blk_mq_free_tag_set(&nbd->tag_set);
 226                 blk_cleanup_disk(disk);
 227         }
 228
 229         /*
 230          * Place this in the last just before the nbd is freed to
 231          * make sure that the disk and the related kobject are also
 232          * totally removed to avoid duplicate creation of the same
 233          * one.
 234          */
 235         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) && nbd->destroy_complete)
 236                 complete(nbd->destroy_complete);
 237
 238         kfree(nbd);
 239 }
 240
 241 static void nbd_put(struct nbd_device *nbd)
 242 {
 243         if (refcount_dec_and_mutex_lock(&nbd->refs,
 244                                         &nbd_index_mutex)) {
 245                 idr_remove(&nbd_index_idr, nbd->index);
 246                 nbd_dev_remove(nbd);
 247                 mutex_unlock(&nbd_index_mutex);
 248         }
 249 }
 250
 251 static int nbd_disconnected(struct nbd_config *config)
 252 {
 253         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 254                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 255 }
 256
 257 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 258                                 int notify)
 259 {
 260         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 261                 struct link_dead_args *args;
 262                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 263                 if (args) {
 264                         INIT_WORK(&args->work, nbd_dead_link_work);
 265                         args->index = nbd->index;
 266                         queue_work(system_wq, &args->work);
 267                 }
 268         }
 269         if (!nsock->dead) {
 270                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 271                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 272                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 273                                                &nbd->config->runtime_flags)) {
 274                                 set_bit(NBD_RT_DISCONNECTED,
 275                                         &nbd->config->runtime_flags);
 276                                 dev_info(nbd_to_dev(nbd),
 277                                         "Disconnected due to user request.\n");
 278                         }
 279                 }
 280         }
 281         nsock->dead = true;
 282         nsock->pending = NULL;
 283         nsock->sent = 0;
 284 }
 285
 286 static void nbd_size_clear(struct nbd_device *nbd)
 287 {
 288         if (nbd->config->bytesize) {
 289                 set_capacity(nbd->disk, 0);
 290                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 291         }
 292 }
 293
 294 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 295                 loff_t blksize)
 296 {
 297         if (!blksize)
 298                 blksize = NBD_DEF_BLKSIZE;
 299         if (blksize < 512 || blksize > PAGE_SIZE || !is_power_of_2(blksize))
 300                 return -EINVAL;
 301
 302         nbd->config->bytesize = bytesize;
 303         nbd->config->blksize = blksize;
 304
 305         if (!nbd->task_recv)
 306                 return 0;
 307
 308         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 309                 nbd->disk->queue->limits.discard_granularity = blksize;
 310                 nbd->disk->queue->limits.discard_alignment = blksize;
 311                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 312         }
 313         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 314         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 315
 316         if (max_part)
 317                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 318         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 319                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 320         return 0;
 321 }
 322
 323 static void nbd_complete_rq(struct request *req)
 324 {
 325         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 326
 327         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 328                 cmd->status ? "failed" : "done");
 329
 330         blk_mq_end_request(req, cmd->status);
 331 }
 332
 333 /*
 334  * Forcibly shutdown the socket causing all listeners to error
 335  */
 336 static void sock_shutdown(struct nbd_device *nbd)
 337 {
 338         struct nbd_config *config = nbd->config;
 339         int i;
 340
 341         if (config->num_connections == 0)
 342                 return;
 343         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 344                 return;
 345
 346         for (i = 0; i < config->num_connections; i++) {
 347                 struct nbd_sock *nsock = config->socks[i];
 348                 mutex_lock(&nsock->tx_lock);
 349                 nbd_mark_nsock_dead(nbd, nsock, 0);
 350                 mutex_unlock(&nsock->tx_lock);
 351         }
 352         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 353 }
 354
 355 static u32 req_to_nbd_cmd_type(struct request *req)
 356 {
 357         switch (req_op(req)) {
 358         case REQ_OP_DISCARD:
 359                 return NBD_CMD_TRIM;
 360         case REQ_OP_FLUSH:
 361                 return NBD_CMD_FLUSH;
 362         case REQ_OP_WRITE:
 363                 return NBD_CMD_WRITE;
 364         case REQ_OP_READ:
 365                 return NBD_CMD_READ;
 366         default:
 367                 return U32_MAX;
 368         }
 369 }
 370
 371 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 372                                                  bool reserved)
 373 {
 374         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 375         struct nbd_device *nbd = cmd->nbd;
 376         struct nbd_config *config;
 377
 378         if (!mutex_trylock(&cmd->lock))
 379                 return BLK_EH_RESET_TIMER;
 380
 381         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 382                 cmd->status = BLK_STS_TIMEOUT;
 383                 mutex_unlock(&cmd->lock);
 384                 goto done;
 385         }
 386         config = nbd->config;
 387
 388         if (config->num_connections > 1 ||
 389             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 390                 dev_err_ratelimited(nbd_to_dev(nbd),
 391                                     "Connection timed out, retrying (%d/%d alive)\n",
 392                                     atomic_read(&config->live_connections),
 393                                     config->num_connections);
 394                 /*
 395                  * Hooray we have more connections, requeue this IO, the submit
 396                  * path will put it on a real connection. Or if only one
 397                  * connection is configured, the submit path will wait util
 398                  * a new connection is reconfigured or util dead timeout.
 399                  */
 400                 if (config->socks) {
 401                         if (cmd->index < config->num_connections) {
 402                                 struct nbd_sock *nsock =
 403                                         config->socks[cmd->index];
 404                                 mutex_lock(&nsock->tx_lock);
 405                                 /* We can have multiple outstanding requests, so
 406                                  * we don't want to mark the nsock dead if we've
 407                                  * already reconnected with a new socket, so
 408                                  * only mark it dead if its the same socket we
 409                                  * were sent out on.
 410                                  */
 411                                 if (cmd->cookie == nsock->cookie)
 412                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 413                                 mutex_unlock(&nsock->tx_lock);
 414                         }
 415                         mutex_unlock(&cmd->lock);
 416                         nbd_requeue_cmd(cmd);
 417                         nbd_config_put(nbd);
 418                         return BLK_EH_DONE;
 419                 }
 420         }
 421
 422         if (!nbd->tag_set.timeout) {
 423                 /*
 424                  * Userspace sets timeout=0 to disable socket disconnection,
 425                  * so just warn and reset the timer.
 426                  */
 427                 struct nbd_sock *nsock = config->socks[cmd->index];
 428                 cmd->retries++;
 429                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 430                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 431                         (unsigned long long)blk_rq_pos(req) << 9,
 432                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 433
 434                 mutex_lock(&nsock->tx_lock);
 435                 if (cmd->cookie != nsock->cookie) {
 436                         nbd_requeue_cmd(cmd);
 437                         mutex_unlock(&nsock->tx_lock);
 438                         mutex_unlock(&cmd->lock);
 439                         nbd_config_put(nbd);
 440                         return BLK_EH_DONE;
 441                 }
 442                 mutex_unlock(&nsock->tx_lock);
 443                 mutex_unlock(&cmd->lock);
 444                 nbd_config_put(nbd);
 445                 return BLK_EH_RESET_TIMER;
 446         }
 447
 448         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 449         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 450         cmd->status = BLK_STS_IOERR;
 451         mutex_unlock(&cmd->lock);
 452         sock_shutdown(nbd);
 453         nbd_config_put(nbd);
 454 done:
 455         blk_mq_complete_request(req);
 456         return BLK_EH_DONE;
 457 }
 458
 459 /*
 460  *  Send or receive packet.
 461  */
 462 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 463                      struct iov_iter *iter, int msg_flags, int *sent)
 464 {
 465         struct nbd_config *config = nbd->config;
 466         struct socket *sock = config->socks[index]->sock;
 467         int result;
 468         struct msghdr msg;
 469         unsigned int noreclaim_flag;
 470
 471         if (unlikely(!sock)) {
 472                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 473                         "Attempted %s on closed socket in sock_xmit\n",
 474                         (send ? "send" : "recv"));
 475                 return -EINVAL;
 476         }
 477
 478         msg.msg_iter = *iter;
 479
 480         noreclaim_flag = memalloc_noreclaim_save();
 481         do {
 482                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 483                 msg.msg_name = NULL;
 484                 msg.msg_namelen = 0;
 485                 msg.msg_control = NULL;
 486                 msg.msg_controllen = 0;
 487                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 488
 489                 if (send)
 490                         result = sock_sendmsg(sock, &msg);
 491                 else
 492                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 493
 494                 if (result <= 0) {
 495                         if (result == 0)
 496                                 result = -EPIPE; /* short read */
 497                         break;
 498                 }
 499                 if (sent)
 500                         *sent += result;
 501         } while (msg_data_left(&msg));
 502
 503         memalloc_noreclaim_restore(noreclaim_flag);
 504
 505         return result;
 506 }
 507
 508 /*
 509  * Different settings for sk->sk_sndtimeo can result in different return values
 510  * if there is a signal pending when we enter sendmsg, because reasons?
 511  */
 512 static inline int was_interrupted(int result)
 513 {
 514         return result == -ERESTARTSYS || result == -EINTR;
 515 }
 516
 517 /* always call with the tx_lock held */
 518 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 519 {
 520         struct request *req = blk_mq_rq_from_pdu(cmd);
 521         struct nbd_config *config = nbd->config;
 522         struct nbd_sock *nsock = config->socks[index];
 523         int result;
 524         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 525         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 526         struct iov_iter from;
 527         unsigned long size = blk_rq_bytes(req);
 528         struct bio *bio;
 529         u64 handle;
 530         u32 type;
 531         u32 nbd_cmd_flags = 0;
 532         int sent = nsock->sent, skip = 0;
 533
 534         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 535
 536         type = req_to_nbd_cmd_type(req);
 537         if (type == U32_MAX)
 538                 return -EIO;
 539
 540         if (rq_data_dir(req) == WRITE &&
 541             (config->flags & NBD_FLAG_READ_ONLY)) {
 542                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 543                                     "Write on read-only\n");
 544                 return -EIO;
 545         }
 546
 547         if (req->cmd_flags & REQ_FUA)
 548                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 549
 550         /* We did a partial send previously, and we at least sent the whole
 551          * request struct, so just go and send the rest of the pages in the
 552          * request.
 553          */
 554         if (sent) {
 555                 if (sent >= sizeof(request)) {
 556                         skip = sent - sizeof(request);
 557
 558                         /* initialize handle for tracing purposes */
 559                         handle = nbd_cmd_handle(cmd);
 560
 561                         goto send_pages;
 562                 }
 563                 iov_iter_advance(&from, sent);
 564         } else {
 565                 cmd->cmd_cookie++;
 566         }
 567         cmd->index = index;
 568         cmd->cookie = nsock->cookie;
 569         cmd->retries = 0;
 570         request.type = htonl(type | nbd_cmd_flags);
 571         if (type != NBD_CMD_FLUSH) {
 572                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 573                 request.len = htonl(size);
 574         }
 575         handle = nbd_cmd_handle(cmd);
 576         memcpy(request.handle, &handle, sizeof(handle));
 577
 578         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 579
 580         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 581                 req, nbdcmd_to_ascii(type),
 582                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 583         result = sock_xmit(nbd, index, 1, &from,
 584                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 585         trace_nbd_header_sent(req, handle);
 586         if (result <= 0) {
 587                 if (was_interrupted(result)) {
 588                         /* If we havne't sent anything we can just return BUSY,
 589                          * however if we have sent something we need to make
 590                          * sure we only allow this req to be sent until we are
 591                          * completely done.
 592                          */
 593                         if (sent) {
 594                                 nsock->pending = req;
 595                                 nsock->sent = sent;
 596                         }
 597                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 598                         return BLK_STS_RESOURCE;
 599                 }
 600                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 601                         "Send control failed (result %d)\n", result);
 602                 return -EAGAIN;
 603         }
 604 send_pages:
 605         if (type != NBD_CMD_WRITE)
 606                 goto out;
 607
 608         bio = req->bio;
 609         while (bio) {
 610                 struct bio *next = bio->bi_next;
 611                 struct bvec_iter iter;
 612                 struct bio_vec bvec;
 613
 614                 bio_for_each_segment(bvec, bio, iter) {
 615                         bool is_last = !next && bio_iter_last(bvec, iter);
 616                         int flags = is_last ? 0 : MSG_MORE;
 617
 618                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 619                                 req, bvec.bv_len);
 620                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 621                         if (skip) {
 622                                 if (skip >= iov_iter_count(&from)) {
 623                                         skip -= iov_iter_count(&from);
 624                                         continue;
 625                                 }
 626                                 iov_iter_advance(&from, skip);
 627                                 skip = 0;
 628                         }
 629                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 630                         if (result <= 0) {
 631                                 if (was_interrupted(result)) {
 632                                         /* We've already sent the header, we
 633                                          * have no choice but to set pending and
 634                                          * return BUSY.
 635                                          */
 636                                         nsock->pending = req;
 637                                         nsock->sent = sent;
 638                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 639                                         return BLK_STS_RESOURCE;
 640                                 }
 641                                 dev_err(disk_to_dev(nbd->disk),
 642                                         "Send data failed (result %d)\n",
 643                                         result);
 644                                 return -EAGAIN;
 645                         }
 646                         /*
 647                          * The completion might already have come in,
 648                          * so break for the last one instead of letting
 649                          * the iterator do it. This prevents use-after-free
 650                          * of the bio.
 651                          */
 652                         if (is_last)
 653                                 break;
 654                 }
 655                 bio = next;
 656         }
 657 out:
 658         trace_nbd_payload_sent(req, handle);
 659         nsock->pending = NULL;
 660         nsock->sent = 0;
 661         return 0;
 662 }
 663
 664 /* NULL returned = something went wrong, inform userspace */
 665 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 666 {
 667         struct nbd_config *config = nbd->config;
 668         int result;
 669         struct nbd_reply reply;
 670         struct nbd_cmd *cmd;
 671         struct request *req = NULL;
 672         u64 handle;
 673         u16 hwq;
 674         u32 tag;
 675         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 676         struct iov_iter to;
 677         int ret = 0;
 678
 679         reply.magic = 0;
 680         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 681         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 682         if (result <= 0) {
 683                 if (!nbd_disconnected(config))
 684                         dev_err(disk_to_dev(nbd->disk),
 685                                 "Receive control failed (result %d)\n", result);
 686                 return ERR_PTR(result);
 687         }
 688
 689         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 690                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 691                                 (unsigned long)ntohl(reply.magic));
 692                 return ERR_PTR(-EPROTO);
 693         }
 694
 695         memcpy(&handle, reply.handle, sizeof(handle));
 696         tag = nbd_handle_to_tag(handle);
 697         hwq = blk_mq_unique_tag_to_hwq(tag);
 698         if (hwq < nbd->tag_set.nr_hw_queues)
 699                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 700                                        blk_mq_unique_tag_to_tag(tag));
 701         if (!req || !blk_mq_request_started(req)) {
 702                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 703                         tag, req);
 704                 return ERR_PTR(-ENOENT);
 705         }
 706         trace_nbd_header_received(req, handle);
 707         cmd = blk_mq_rq_to_pdu(req);
 708
 709         mutex_lock(&cmd->lock);
 710         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 711                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 712                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 713                 ret = -ENOENT;
 714                 goto out;
 715         }
 716         if (cmd->status != BLK_STS_OK) {
 717                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 718                         req);
 719                 ret = -ENOENT;
 720                 goto out;
 721         }
 722         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 723                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 724                         req);
 725                 ret = -ENOENT;
 726                 goto out;
 727         }
 728         if (ntohl(reply.error)) {
 729                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 730                         ntohl(reply.error));
 731                 cmd->status = BLK_STS_IOERR;
 732                 goto out;
 733         }
 734
 735         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 736         if (rq_data_dir(req) != WRITE) {
 737                 struct req_iterator iter;
 738                 struct bio_vec bvec;
 739
 740                 rq_for_each_segment(bvec, req, iter) {
 741                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 742                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 743                         if (result <= 0) {
 744                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 745                                         result);
 746                                 /*
 747                                  * If we've disconnected, we need to make sure we
 748                                  * complete this request, otherwise error out
 749                                  * and let the timeout stuff handle resubmitting
 750                                  * this request onto another connection.
 751                                  */
 752                                 if (nbd_disconnected(config)) {
 753                                         cmd->status = BLK_STS_IOERR;
 754                                         goto out;
 755                                 }
 756                                 ret = -EIO;
 757                                 goto out;
 758                         }
 759                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 760                                 req, bvec.bv_len);
 761                 }
 762         }
 763 out:
 764         trace_nbd_payload_received(req, handle);
 765         mutex_unlock(&cmd->lock);
 766         return ret ? ERR_PTR(ret) : cmd;
 767 }
 768
 769 static void recv_work(struct work_struct *work)
 770 {
 771         struct recv_thread_args *args = container_of(work,
 772                                                      struct recv_thread_args,
 773                                                      work);
 774         struct nbd_device *nbd = args->nbd;
 775         struct nbd_config *config = nbd->config;
 776         struct nbd_cmd *cmd;
 777         struct request *rq;
 778
 779         while (1) {
 780                 cmd = nbd_read_stat(nbd, args->index);
 781                 if (IS_ERR(cmd)) {
 782                         struct nbd_sock *nsock = config->socks[args->index];
 783
 784                         mutex_lock(&nsock->tx_lock);
 785                         nbd_mark_nsock_dead(nbd, nsock, 1);
 786                         mutex_unlock(&nsock->tx_lock);
 787                         break;
 788                 }
 789
 790                 rq = blk_mq_rq_from_pdu(cmd);
 791                 if (likely(!blk_should_fake_timeout(rq->q)))
 792                         blk_mq_complete_request(rq);
 793         }
 794         nbd_config_put(nbd);
 795         atomic_dec(&config->recv_threads);
 796         wake_up(&config->recv_wq);
 797         kfree(args);
 798 }
 799
 800 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 801 {
 802         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 803
 804         mutex_lock(&cmd->lock);
 805         cmd->status = BLK_STS_IOERR;
 806         mutex_unlock(&cmd->lock);
 807
 808         blk_mq_complete_request(req);
 809         return true;
 810 }
 811
 812 static void nbd_clear_que(struct nbd_device *nbd)
 813 {
 814         blk_mq_quiesce_queue(nbd->disk->queue);
 815         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 816         blk_mq_unquiesce_queue(nbd->disk->queue);
 817         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 818 }
 819
 820 static int find_fallback(struct nbd_device *nbd, int index)
 821 {
 822         struct nbd_config *config = nbd->config;
 823         int new_index = -1;
 824         struct nbd_sock *nsock = config->socks[index];
 825         int fallback = nsock->fallback_index;
 826
 827         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 828                 return new_index;
 829
 830         if (config->num_connections <= 1) {
 831                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 832                                     "Dead connection, failed to find a fallback\n");
 833                 return new_index;
 834         }
 835
 836         if (fallback >= 0 && fallback < config->num_connections &&
 837             !config->socks[fallback]->dead)
 838                 return fallback;
 839
 840         if (nsock->fallback_index < 0 ||
 841             nsock->fallback_index >= config->num_connections ||
 842             config->socks[nsock->fallback_index]->dead) {
 843                 int i;
 844                 for (i = 0; i < config->num_connections; i++) {
 845                         if (i == index)
 846                                 continue;
 847                         if (!config->socks[i]->dead) {
 848                                 new_index = i;
 849                                 break;
 850                         }
 851                 }
 852                 nsock->fallback_index = new_index;
 853                 if (new_index < 0) {
 854                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 855                                             "Dead connection, failed to find a fallback\n");
 856                         return new_index;
 857                 }
 858         }
 859         new_index = nsock->fallback_index;
 860         return new_index;
 861 }
 862
 863 static int wait_for_reconnect(struct nbd_device *nbd)
 864 {
 865         struct nbd_config *config = nbd->config;
 866         if (!config->dead_conn_timeout)
 867                 return 0;
 868         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 869                 return 0;
 870         return wait_event_timeout(config->conn_wait,
 871                                   atomic_read(&config->live_connections) > 0,
 872                                   config->dead_conn_timeout) > 0;
 873 }
 874
 875 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 876 {
 877         struct request *req = blk_mq_rq_from_pdu(cmd);
 878         struct nbd_device *nbd = cmd->nbd;
 879         struct nbd_config *config;
 880         struct nbd_sock *nsock;
 881         int ret;
 882
 883         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 884                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 885                                     "Socks array is empty\n");
 886                 blk_mq_start_request(req);
 887                 return -EINVAL;
 888         }
 889         config = nbd->config;
 890
 891         if (index >= config->num_connections) {
 892                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 893                                     "Attempted send on invalid socket\n");
 894                 nbd_config_put(nbd);
 895                 blk_mq_start_request(req);
 896                 return -EINVAL;
 897         }
 898         cmd->status = BLK_STS_OK;
 899 again:
 900         nsock = config->socks[index];
 901         mutex_lock(&nsock->tx_lock);
 902         if (nsock->dead) {
 903                 int old_index = index;
 904                 index = find_fallback(nbd, index);
 905                 mutex_unlock(&nsock->tx_lock);
 906                 if (index < 0) {
 907                         if (wait_for_reconnect(nbd)) {
 908                                 index = old_index;
 909                                 goto again;
 910                         }
 911                         /* All the sockets should already be down at this point,
 912                          * we just want to make sure that DISCONNECTED is set so
 913                          * any requests that come in that were queue'ed waiting
 914                          * for the reconnect timer don't trigger the timer again
 915                          * and instead just error out.
 916                          */
 917                         sock_shutdown(nbd);
 918                         nbd_config_put(nbd);
 919                         blk_mq_start_request(req);
 920                         return -EIO;
 921                 }
 922                 goto again;
 923         }
 924
 925         /* Handle the case that we have a pending request that was partially
 926          * transmitted that _has_ to be serviced first.  We need to call requeue
 927          * here so that it gets put _after_ the request that is already on the
 928          * dispatch list.
 929          */
 930         blk_mq_start_request(req);
 931         if (unlikely(nsock->pending && nsock->pending != req)) {
 932                 nbd_requeue_cmd(cmd);
 933                 ret = 0;
 934                 goto out;
 935         }
 936         /*
 937          * Some failures are related to the link going down, so anything that
 938          * returns EAGAIN can be retried on a different socket.
 939          */
 940         ret = nbd_send_cmd(nbd, cmd, index);
 941         if (ret == -EAGAIN) {
 942                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 943                                     "Request send failed, requeueing\n");
 944                 nbd_mark_nsock_dead(nbd, nsock, 1);
 945                 nbd_requeue_cmd(cmd);
 946                 ret = 0;
 947         }
 948 out:
 949         mutex_unlock(&nsock->tx_lock);
 950         nbd_config_put(nbd);
 951         return ret;
 952 }
 953
 954 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 955                         const struct blk_mq_queue_data *bd)
 956 {
 957         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 958         int ret;
 959
 960         /*
 961          * Since we look at the bio's to send the request over the network we
 962          * need to make sure the completion work doesn't mark this request done
 963          * before we are done doing our send.  This keeps us from dereferencing
 964          * freed data if we have particularly fast completions (ie we get the
 965          * completion before we exit sock_xmit on the last bvec) or in the case
 966          * that the server is misbehaving (or there was an error) before we're
 967          * done sending everything over the wire.
 968          */
 969         mutex_lock(&cmd->lock);
 970         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 971
 972         /* We can be called directly from the user space process, which means we
 973          * could possibly have signals pending so our sendmsg will fail.  In
 974          * this case we need to return that we are busy, otherwise error out as
 975          * appropriate.
 976          */
 977         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 978         if (ret < 0)
 979                 ret = BLK_STS_IOERR;
 980         else if (!ret)
 981                 ret = BLK_STS_OK;
 982         mutex_unlock(&cmd->lock);
 983
 984         return ret;
 985 }
 986
 987 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
 988                                      int *err)
 989 {
 990         struct socket *sock;
 991
 992         *err = 0;
 993         sock = sockfd_lookup(fd, err);
 994         if (!sock)
 995                 return NULL;
 996
 997         if (sock->ops->shutdown == sock_no_shutdown) {
 998                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
 999                 *err = -EINVAL;
1000                 sockfd_put(sock);
1001                 return NULL;
1002         }
1003
1004         return sock;
1005 }
1006
1007 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1008                           bool netlink)
1009 {
1010         struct nbd_config *config = nbd->config;
1011         struct socket *sock;
1012         struct nbd_sock **socks;
1013         struct nbd_sock *nsock;
1014         int err;
1015
1016         sock = nbd_get_socket(nbd, arg, &err);
1017         if (!sock)
1018                 return err;
1019
1020         /*
1021          * We need to make sure we don't get any errant requests while we're
1022          * reallocating the ->socks array.
1023          */
1024         blk_mq_freeze_queue(nbd->disk->queue);
1025
1026         if (!netlink && !nbd->task_setup &&
1027             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1028                 nbd->task_setup = current;
1029
1030         if (!netlink &&
1031             (nbd->task_setup != current ||
1032              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1033                 dev_err(disk_to_dev(nbd->disk),
1034                         "Device being setup by another task");
1035                 err = -EBUSY;
1036                 goto put_socket;
1037         }
1038
1039         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1040         if (!nsock) {
1041                 err = -ENOMEM;
1042                 goto put_socket;
1043         }
1044
1045         socks = krealloc(config->socks, (config->num_connections + 1) *
1046                          sizeof(struct nbd_sock *), GFP_KERNEL);
1047         if (!socks) {
1048                 kfree(nsock);
1049                 err = -ENOMEM;
1050                 goto put_socket;
1051         }
1052
1053         config->socks = socks;
1054
1055         nsock->fallback_index = -1;
1056         nsock->dead = false;
1057         mutex_init(&nsock->tx_lock);
1058         nsock->sock = sock;
1059         nsock->pending = NULL;
1060         nsock->sent = 0;
1061         nsock->cookie = 0;
1062         socks[config->num_connections++] = nsock;
1063         atomic_inc(&config->live_connections);
1064         blk_mq_unfreeze_queue(nbd->disk->queue);
1065
1066         return 0;
1067
1068 put_socket:
1069         blk_mq_unfreeze_queue(nbd->disk->queue);
1070         sockfd_put(sock);
1071         return err;
1072 }
1073
1074 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1075 {
1076         struct nbd_config *config = nbd->config;
1077         struct socket *sock, *old;
1078         struct recv_thread_args *args;
1079         int i;
1080         int err;
1081
1082         sock = nbd_get_socket(nbd, arg, &err);
1083         if (!sock)
1084                 return err;
1085
1086         args = kzalloc(sizeof(*args), GFP_KERNEL);
1087         if (!args) {
1088                 sockfd_put(sock);
1089                 return -ENOMEM;
1090         }
1091
1092         for (i = 0; i < config->num_connections; i++) {
1093                 struct nbd_sock *nsock = config->socks[i];
1094
1095                 if (!nsock->dead)
1096                         continue;
1097
1098                 mutex_lock(&nsock->tx_lock);
1099                 if (!nsock->dead) {
1100                         mutex_unlock(&nsock->tx_lock);
1101                         continue;
1102                 }
1103                 sk_set_memalloc(sock->sk);
1104                 if (nbd->tag_set.timeout)
1105                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1106                 atomic_inc(&config->recv_threads);
1107                 refcount_inc(&nbd->config_refs);
1108                 old = nsock->sock;
1109                 nsock->fallback_index = -1;
1110                 nsock->sock = sock;
1111                 nsock->dead = false;
1112                 INIT_WORK(&args->work, recv_work);
1113                 args->index = i;
1114                 args->nbd = nbd;
1115                 nsock->cookie++;
1116                 mutex_unlock(&nsock->tx_lock);
1117                 sockfd_put(old);
1118
1119                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1120
1121                 /* We take the tx_mutex in an error path in the recv_work, so we
1122                  * need to queue_work outside of the tx_mutex.
1123                  */
1124                 queue_work(nbd->recv_workq, &args->work);
1125
1126                 atomic_inc(&config->live_connections);
1127                 wake_up(&config->conn_wait);
1128                 return 0;
1129         }
1130         sockfd_put(sock);
1131         kfree(args);
1132         return -ENOSPC;
1133 }
1134
1135 static void nbd_bdev_reset(struct block_device *bdev)
1136 {
1137         if (bdev->bd_openers > 1)
1138                 return;
1139         set_capacity(bdev->bd_disk, 0);
1140 }
1141
1142 static void nbd_parse_flags(struct nbd_device *nbd)
1143 {
1144         struct nbd_config *config = nbd->config;
1145         if (config->flags & NBD_FLAG_READ_ONLY)
1146                 set_disk_ro(nbd->disk, true);
1147         else
1148                 set_disk_ro(nbd->disk, false);
1149         if (config->flags & NBD_FLAG_SEND_TRIM)
1150                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1151         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1152                 if (config->flags & NBD_FLAG_SEND_FUA)
1153                         blk_queue_write_cache(nbd->disk->queue, true, true);
1154                 else
1155                         blk_queue_write_cache(nbd->disk->queue, true, false);
1156         }
1157         else
1158                 blk_queue_write_cache(nbd->disk->queue, false, false);
1159 }
1160
1161 static void send_disconnects(struct nbd_device *nbd)
1162 {
1163         struct nbd_config *config = nbd->config;
1164         struct nbd_request request = {
1165                 .magic = htonl(NBD_REQUEST_MAGIC),
1166                 .type = htonl(NBD_CMD_DISC),
1167         };
1168         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1169         struct iov_iter from;
1170         int i, ret;
1171
1172         for (i = 0; i < config->num_connections; i++) {
1173                 struct nbd_sock *nsock = config->socks[i];
1174
1175                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1176                 mutex_lock(&nsock->tx_lock);
1177                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1178                 if (ret <= 0)
1179                         dev_err(disk_to_dev(nbd->disk),
1180                                 "Send disconnect failed %d\n", ret);
1181                 mutex_unlock(&nsock->tx_lock);
1182         }
1183 }
1184
1185 static int nbd_disconnect(struct nbd_device *nbd)
1186 {
1187         struct nbd_config *config = nbd->config;
1188
1189         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1190         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1191         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1192         send_disconnects(nbd);
1193         return 0;
1194 }
1195
1196 static void nbd_clear_sock(struct nbd_device *nbd)
1197 {
1198         sock_shutdown(nbd);
1199         nbd_clear_que(nbd);
1200         nbd->task_setup = NULL;
1201 }
1202
1203 static void nbd_config_put(struct nbd_device *nbd)
1204 {
1205         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1206                                         &nbd->config_lock)) {
1207                 struct nbd_config *config = nbd->config;
1208                 nbd_dev_dbg_close(nbd);
1209                 nbd_size_clear(nbd);
1210                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1211                                        &config->runtime_flags))
1212                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1213                 nbd->task_recv = NULL;
1214                 nbd_clear_sock(nbd);
1215                 if (config->num_connections) {
1216                         int i;
1217                         for (i = 0; i < config->num_connections; i++) {
1218                                 sockfd_put(config->socks[i]->sock);
1219                                 kfree(config->socks[i]);
1220                         }
1221                         kfree(config->socks);
1222                 }
1223                 kfree(nbd->config);
1224                 nbd->config = NULL;
1225
1226                 if (nbd->recv_workq)
1227                         destroy_workqueue(nbd->recv_workq);
1228                 nbd->recv_workq = NULL;
1229
1230                 nbd->tag_set.timeout = 0;
1231                 nbd->disk->queue->limits.discard_granularity = 0;
1232                 nbd->disk->queue->limits.discard_alignment = 0;
1233                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1234                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1235
1236                 mutex_unlock(&nbd->config_lock);
1237                 nbd_put(nbd);
1238                 module_put(THIS_MODULE);
1239         }
1240 }
1241
1242 static int nbd_start_device(struct nbd_device *nbd)
1243 {
1244         struct nbd_config *config = nbd->config;
1245         int num_connections = config->num_connections;
1246         int error = 0, i;
1247
1248         if (nbd->task_recv)
1249                 return -EBUSY;
1250         if (!config->socks)
1251                 return -EINVAL;
1252         if (num_connections > 1 &&
1253             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1254                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1255                 return -EINVAL;
1256         }
1257
1258         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1259                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1260                                           WQ_UNBOUND, 0, nbd->index);
1261         if (!nbd->recv_workq) {
1262                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1263                 return -ENOMEM;
1264         }
1265
1266         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1267         nbd->task_recv = current;
1268
1269         nbd_parse_flags(nbd);
1270
1271         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1272         if (error) {
1273                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1274                 return error;
1275         }
1276         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1277
1278         nbd_dev_dbg_init(nbd);
1279         for (i = 0; i < num_connections; i++) {
1280                 struct recv_thread_args *args;
1281
1282                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1283                 if (!args) {
1284                         sock_shutdown(nbd);
1285                         /*
1286                          * If num_connections is m (2 < m),
1287                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1288                          * But NO.(n + 1) failed. We still have n recv threads.
1289                          * So, add flush_workqueue here to prevent recv threads
1290                          * dropping the last config_refs and trying to destroy
1291                          * the workqueue from inside the workqueue.
1292                          */
1293                         if (i)
1294                                 flush_workqueue(nbd->recv_workq);
1295                         return -ENOMEM;
1296                 }
1297                 sk_set_memalloc(config->socks[i]->sock->sk);
1298                 if (nbd->tag_set.timeout)
1299                         config->socks[i]->sock->sk->sk_sndtimeo =
1300                                 nbd->tag_set.timeout;
1301                 atomic_inc(&config->recv_threads);
1302                 refcount_inc(&nbd->config_refs);
1303                 INIT_WORK(&args->work, recv_work);
1304                 args->nbd = nbd;
1305                 args->index = i;
1306                 queue_work(nbd->recv_workq, &args->work);
1307         }
1308         return nbd_set_size(nbd, config->bytesize, config->blksize);
1309 }
1310
1311 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1312 {
1313         struct nbd_config *config = nbd->config;
1314         int ret;
1315
1316         ret = nbd_start_device(nbd);
1317         if (ret)
1318                 return ret;
1319
1320         if (max_part)
1321                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1322         mutex_unlock(&nbd->config_lock);
1323         ret = wait_event_interruptible(config->recv_wq,
1324                                          atomic_read(&config->recv_threads) == 0);
1325         if (ret)
1326                 sock_shutdown(nbd);
1327         flush_workqueue(nbd->recv_workq);
1328
1329         mutex_lock(&nbd->config_lock);
1330         nbd_bdev_reset(bdev);
1331         /* user requested, ignore socket errors */
1332         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1333                 ret = 0;
1334         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1335                 ret = -ETIMEDOUT;
1336         return ret;
1337 }
1338
1339 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1340                                  struct block_device *bdev)
1341 {
1342         sock_shutdown(nbd);
1343         __invalidate_device(bdev, true);
1344         nbd_bdev_reset(bdev);
1345         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1346                                &nbd->config->runtime_flags))
1347                 nbd_config_put(nbd);
1348 }
1349
1350 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1351 {
1352         nbd->tag_set.timeout = timeout * HZ;
1353         if (timeout)
1354                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1355         else
1356                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1357 }
1358
1359 /* Must be called with config_lock held */
1360 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1361                        unsigned int cmd, unsigned long arg)
1362 {
1363         struct nbd_config *config = nbd->config;
1364
1365         switch (cmd) {
1366         case NBD_DISCONNECT:
1367                 return nbd_disconnect(nbd);
1368         case NBD_CLEAR_SOCK:
1369                 nbd_clear_sock_ioctl(nbd, bdev);
1370                 return 0;
1371         case NBD_SET_SOCK:
1372                 return nbd_add_socket(nbd, arg, false);
1373         case NBD_SET_BLKSIZE:
1374                 return nbd_set_size(nbd, config->bytesize, arg);
1375         case NBD_SET_SIZE:
1376                 return nbd_set_size(nbd, arg, config->blksize);
1377         case NBD_SET_SIZE_BLOCKS:
1378                 return nbd_set_size(nbd, arg * config->blksize,
1379                                     config->blksize);
1380         case NBD_SET_TIMEOUT:
1381                 nbd_set_cmd_timeout(nbd, arg);
1382                 return 0;
1383
1384         case NBD_SET_FLAGS:
1385                 config->flags = arg;
1386                 return 0;
1387         case NBD_DO_IT:
1388                 return nbd_start_device_ioctl(nbd, bdev);
1389         case NBD_CLEAR_QUE:
1390                 /*
1391                  * This is for compatibility only.  The queue is always cleared
1392                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1393                  */
1394                 return 0;
1395         case NBD_PRINT_DEBUG:
1396                 /*
1397                  * For compatibility only, we no longer keep a list of
1398                  * outstanding requests.
1399                  */
1400                 return 0;
1401         }
1402         return -ENOTTY;
1403 }
1404
1405 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1406                      unsigned int cmd, unsigned long arg)
1407 {
1408         struct nbd_device *nbd = bdev->bd_disk->private_data;
1409         struct nbd_config *config = nbd->config;
1410         int error = -EINVAL;
1411
1412         if (!capable(CAP_SYS_ADMIN))
1413                 return -EPERM;
1414
1415         /* The block layer will pass back some non-nbd ioctls in case we have
1416          * special handling for them, but we don't so just return an error.
1417          */
1418         if (_IOC_TYPE(cmd) != 0xab)
1419                 return -EINVAL;
1420
1421         mutex_lock(&nbd->config_lock);
1422
1423         /* Don't allow ioctl operations on a nbd device that was created with
1424          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1425          */
1426         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1427             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1428                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1429         else
1430                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1431         mutex_unlock(&nbd->config_lock);
1432         return error;
1433 }
1434
1435 static struct nbd_config *nbd_alloc_config(void)
1436 {
1437         struct nbd_config *config;
1438
1439         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1440         if (!config)
1441                 return NULL;
1442         atomic_set(&config->recv_threads, 0);
1443         init_waitqueue_head(&config->recv_wq);
1444         init_waitqueue_head(&config->conn_wait);
1445         config->blksize = NBD_DEF_BLKSIZE;
1446         atomic_set(&config->live_connections, 0);
1447         try_module_get(THIS_MODULE);
1448         return config;
1449 }
1450
1451 static int nbd_open(struct block_device *bdev, fmode_t mode)
1452 {
1453         struct nbd_device *nbd;
1454         int ret = 0;
1455
1456         mutex_lock(&nbd_index_mutex);
1457         nbd = bdev->bd_disk->private_data;
1458         if (!nbd) {
1459                 ret = -ENXIO;
1460                 goto out;
1461         }
1462         if (!refcount_inc_not_zero(&nbd->refs)) {
1463                 ret = -ENXIO;
1464                 goto out;
1465         }
1466         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1467                 struct nbd_config *config;
1468
1469                 mutex_lock(&nbd->config_lock);
1470                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1471                         mutex_unlock(&nbd->config_lock);
1472                         goto out;
1473                 }
1474                 config = nbd->config = nbd_alloc_config();
1475                 if (!config) {
1476                         ret = -ENOMEM;
1477                         mutex_unlock(&nbd->config_lock);
1478                         goto out;
1479                 }
1480                 refcount_set(&nbd->config_refs, 1);
1481                 refcount_inc(&nbd->refs);
1482                 mutex_unlock(&nbd->config_lock);
1483                 if (max_part)
1484                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1485         } else if (nbd_disconnected(nbd->config)) {
1486                 if (max_part)
1487                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1488         }
1489 out:
1490         mutex_unlock(&nbd_index_mutex);
1491         return ret;
1492 }
1493
1494 static void nbd_release(struct gendisk *disk, fmode_t mode)
1495 {
1496         struct nbd_device *nbd = disk->private_data;
1497
1498         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1499                         disk->part0->bd_openers == 0)
1500                 nbd_disconnect_and_put(nbd);
1501
1502         nbd_config_put(nbd);
1503         nbd_put(nbd);
1504 }
1505
1506 static const struct block_device_operations nbd_fops =
1507 {
1508         .owner =        THIS_MODULE,
1509         .open =         nbd_open,
1510         .release =      nbd_release,
1511         .ioctl =        nbd_ioctl,
1512         .compat_ioctl = nbd_ioctl,
1513 };
1514
1515 #if IS_ENABLED(CONFIG_DEBUG_FS)
1516
1517 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1518 {
1519         struct nbd_device *nbd = s->private;
1520
1521         if (nbd->task_recv)
1522                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1523
1524         return 0;
1525 }
1526
1527 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1528
1529 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1530 {
1531         struct nbd_device *nbd = s->private;
1532         u32 flags = nbd->config->flags;
1533
1534         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1535
1536         seq_puts(s, "Known flags:\n");
1537
1538         if (flags & NBD_FLAG_HAS_FLAGS)
1539                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1540         if (flags & NBD_FLAG_READ_ONLY)
1541                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1542         if (flags & NBD_FLAG_SEND_FLUSH)
1543                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1544         if (flags & NBD_FLAG_SEND_FUA)
1545                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1546         if (flags & NBD_FLAG_SEND_TRIM)
1547                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1548
1549         return 0;
1550 }
1551
1552 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1553
1554 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1555 {
1556         struct dentry *dir;
1557         struct nbd_config *config = nbd->config;
1558
1559         if (!nbd_dbg_dir)
1560                 return -EIO;
1561
1562         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1563         if (!dir) {
1564                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1565                         nbd_name(nbd));
1566                 return -EIO;
1567         }
1568         config->dbg_dir = dir;
1569
1570         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1571         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1572         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1573         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1574         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1575
1576         return 0;
1577 }
1578
1579 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1580 {
1581         debugfs_remove_recursive(nbd->config->dbg_dir);
1582 }
1583
1584 static int nbd_dbg_init(void)
1585 {
1586         struct dentry *dbg_dir;
1587
1588         dbg_dir = debugfs_create_dir("nbd", NULL);
1589         if (!dbg_dir)
1590                 return -EIO;
1591
1592         nbd_dbg_dir = dbg_dir;
1593
1594         return 0;
1595 }
1596
1597 static void nbd_dbg_close(void)
1598 {
1599         debugfs_remove_recursive(nbd_dbg_dir);
1600 }
1601
1602 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1603
1604 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1605 {
1606         return 0;
1607 }
1608
1609 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1610 {
1611 }
1612
1613 static int nbd_dbg_init(void)
1614 {
1615         return 0;
1616 }
1617
1618 static void nbd_dbg_close(void)
1619 {
1620 }
1621
1622 #endif
1623
1624 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1625                             unsigned int hctx_idx, unsigned int numa_node)
1626 {
1627         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1628         cmd->nbd = set->driver_data;
1629         cmd->flags = 0;
1630         mutex_init(&cmd->lock);
1631         return 0;
1632 }
1633
1634 static const struct blk_mq_ops nbd_mq_ops = {
1635         .queue_rq       = nbd_queue_rq,
1636         .complete       = nbd_complete_rq,
1637         .init_request   = nbd_init_request,
1638         .timeout        = nbd_xmit_timeout,
1639 };
1640
1641 static int nbd_dev_add(int index)
1642 {
1643         struct nbd_device *nbd;
1644         struct gendisk *disk;
1645         int err = -ENOMEM;
1646
1647         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1648         if (!nbd)
1649                 goto out;
1650
1651         nbd->tag_set.ops = &nbd_mq_ops;
1652         nbd->tag_set.nr_hw_queues = 1;
1653         nbd->tag_set.queue_depth = 128;
1654         nbd->tag_set.numa_node = NUMA_NO_NODE;
1655         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1656         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1657                 BLK_MQ_F_BLOCKING;
1658         nbd->tag_set.driver_data = nbd;
1659         nbd->destroy_complete = NULL;
1660
1661         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1662         if (err)
1663                 goto out_free_nbd;
1664
1665         if (index >= 0) {
1666                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1667                                 GFP_KERNEL);
1668                 if (err == -ENOSPC)
1669                         err = -EEXIST;
1670         } else {
1671                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1672                 if (err >= 0)
1673                         index = err;
1674         }
1675         if (err < 0)
1676                 goto out_free_tags;
1677         nbd->index = index;
1678
1679         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1680         if (IS_ERR(disk)) {
1681                 err = PTR_ERR(disk);
1682                 goto out_free_idr;
1683         }
1684         nbd->disk = disk;
1685
1686         /*
1687          * Tell the block layer that we are not a rotational device
1688          */
1689         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1690         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1691         disk->queue->limits.discard_granularity = 0;
1692         disk->queue->limits.discard_alignment = 0;
1693         blk_queue_max_discard_sectors(disk->queue, 0);
1694         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1695         blk_queue_max_segments(disk->queue, USHRT_MAX);
1696         blk_queue_max_hw_sectors(disk->queue, 65536);
1697         disk->queue->limits.max_sectors = 256;
1698
1699         mutex_init(&nbd->config_lock);
1700         refcount_set(&nbd->config_refs, 0);
1701         refcount_set(&nbd->refs, 1);
1702         INIT_LIST_HEAD(&nbd->list);
1703         disk->major = NBD_MAJOR;
1704         disk->first_minor = index << part_shift;
1705         disk->minors = 1 << part_shift;
1706         disk->fops = &nbd_fops;
1707         disk->private_data = nbd;
1708         sprintf(disk->disk_name, "nbd%d", index);
1709         add_disk(disk);
1710         nbd_total_devices++;
1711         return index;
1712
1713 out_free_idr:
1714         idr_remove(&nbd_index_idr, index);
1715 out_free_tags:
1716         blk_mq_free_tag_set(&nbd->tag_set);
1717 out_free_nbd:
1718         kfree(nbd);
1719 out:
1720         return err;
1721 }
1722
1723 static int find_free_cb(int id, void *ptr, void *data)
1724 {
1725         struct nbd_device *nbd = ptr;
1726         struct nbd_device **found = data;
1727
1728         if (!refcount_read(&nbd->config_refs)) {
1729                 *found = nbd;
1730                 return 1;
1731         }
1732         return 0;
1733 }
1734
1735 /* Netlink interface. */
1736 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1737         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1738         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1739         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1740         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1741         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1742         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1743         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1744         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1745         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1746 };
1747
1748 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1749         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1750 };
1751
1752 /* We don't use this right now since we don't parse the incoming list, but we
1753  * still want it here so userspace knows what to expect.
1754  */
1755 static const struct nla_policy __attribute__((unused))
1756 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1757         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1758         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1759 };
1760
1761 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1762 {
1763         struct nbd_config *config = nbd->config;
1764         u64 bsize = config->blksize;
1765         u64 bytes = config->bytesize;
1766
1767         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1768                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1769
1770         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1771                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1772
1773         if (bytes != config->bytesize || bsize != config->blksize)
1774                 return nbd_set_size(nbd, bytes, bsize);
1775         return 0;
1776 }
1777
1778 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1779 {
1780         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1781         struct nbd_device *nbd = NULL;
1782         struct nbd_config *config;
1783         int index = -1;
1784         int ret;
1785         bool put_dev = false;
1786
1787         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1788                 return -EPERM;
1789
1790         if (info->attrs[NBD_ATTR_INDEX])
1791                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1792         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1793                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1794                 return -EINVAL;
1795         }
1796         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1797                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1798                 return -EINVAL;
1799         }
1800 again:
1801         mutex_lock(&nbd_index_mutex);
1802         if (index == -1) {
1803                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1804                 if (ret == 0) {
1805                         int new_index;
1806                         new_index = nbd_dev_add(-1);
1807                         if (new_index < 0) {
1808                                 mutex_unlock(&nbd_index_mutex);
1809                                 printk(KERN_ERR "nbd: failed to add new device\n");
1810                                 return new_index;
1811                         }
1812                         nbd = idr_find(&nbd_index_idr, new_index);
1813                 }
1814         } else {
1815                 nbd = idr_find(&nbd_index_idr, index);
1816                 if (!nbd) {
1817                         ret = nbd_dev_add(index);
1818                         if (ret < 0) {
1819                                 mutex_unlock(&nbd_index_mutex);
1820                                 printk(KERN_ERR "nbd: failed to add new device\n");
1821                                 return ret;
1822                         }
1823                         nbd = idr_find(&nbd_index_idr, index);
1824                 }
1825         }
1826         if (!nbd) {
1827                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1828                        index);
1829                 mutex_unlock(&nbd_index_mutex);
1830                 return -EINVAL;
1831         }
1832
1833         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1834             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1835                 nbd->destroy_complete = &destroy_complete;
1836                 mutex_unlock(&nbd_index_mutex);
1837
1838                 /* Wait untill the the nbd stuff is totally destroyed */
1839                 wait_for_completion(&destroy_complete);
1840                 goto again;
1841         }
1842
1843         if (!refcount_inc_not_zero(&nbd->refs)) {
1844                 mutex_unlock(&nbd_index_mutex);
1845                 if (index == -1)
1846                         goto again;
1847                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1848                        index);
1849                 return -EINVAL;
1850         }
1851         mutex_unlock(&nbd_index_mutex);
1852
1853         mutex_lock(&nbd->config_lock);
1854         if (refcount_read(&nbd->config_refs)) {
1855                 mutex_unlock(&nbd->config_lock);
1856                 nbd_put(nbd);
1857                 if (index == -1)
1858                         goto again;
1859                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1860                 return -EBUSY;
1861         }
1862         if (WARN_ON(nbd->config)) {
1863                 mutex_unlock(&nbd->config_lock);
1864                 nbd_put(nbd);
1865                 return -EINVAL;
1866         }
1867         config = nbd->config = nbd_alloc_config();
1868         if (!nbd->config) {
1869                 mutex_unlock(&nbd->config_lock);
1870                 nbd_put(nbd);
1871                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1872                 return -ENOMEM;
1873         }
1874         refcount_set(&nbd->config_refs, 1);
1875         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1876
1877         ret = nbd_genl_size_set(info, nbd);
1878         if (ret)
1879                 goto out;
1880
1881         if (info->attrs[NBD_ATTR_TIMEOUT])
1882                 nbd_set_cmd_timeout(nbd,
1883                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1884         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1885                 config->dead_conn_timeout =
1886                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1887                 config->dead_conn_timeout *= HZ;
1888         }
1889         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1890                 config->flags =
1891                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1892         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1893                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1894                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1895                         /*
1896                          * We have 1 ref to keep the device around, and then 1
1897                          * ref for our current operation here, which will be
1898                          * inherited by the config.  If we already have
1899                          * DESTROY_ON_DISCONNECT set then we know we don't have
1900                          * that extra ref already held so we don't need the
1901                          * put_dev.
1902                          */
1903                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1904                                               &nbd->flags))
1905                                 put_dev = true;
1906                 } else {
1907                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1908                                                &nbd->flags))
1909                                 refcount_inc(&nbd->refs);
1910                 }
1911                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1912                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1913                                 &config->runtime_flags);
1914                 }
1915         }
1916
1917         if (info->attrs[NBD_ATTR_SOCKETS]) {
1918                 struct nlattr *attr;
1919                 int rem, fd;
1920
1921                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1922                                     rem) {
1923                         struct nlattr *socks[NBD_SOCK_MAX+1];
1924
1925                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1926                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1927                                 ret = -EINVAL;
1928                                 goto out;
1929                         }
1930                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1931                                                           attr,
1932                                                           nbd_sock_policy,
1933                                                           info->extack);
1934                         if (ret != 0) {
1935                                 printk(KERN_ERR "nbd: error processing sock list\n");
1936                                 ret = -EINVAL;
1937                                 goto out;
1938                         }
1939                         if (!socks[NBD_SOCK_FD])
1940                                 continue;
1941                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1942                         ret = nbd_add_socket(nbd, fd, true);
1943                         if (ret)
1944                                 goto out;
1945                 }
1946         }
1947         ret = nbd_start_device(nbd);
1948 out:
1949         mutex_unlock(&nbd->config_lock);
1950         if (!ret) {
1951                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
1952                 refcount_inc(&nbd->config_refs);
1953                 nbd_connect_reply(info, nbd->index);
1954         }
1955         nbd_config_put(nbd);
1956         if (put_dev)
1957                 nbd_put(nbd);
1958         return ret;
1959 }
1960
1961 static void nbd_disconnect_and_put(struct nbd_device *nbd)
1962 {
1963         mutex_lock(&nbd->config_lock);
1964         nbd_disconnect(nbd);
1965         nbd_clear_sock(nbd);
1966         mutex_unlock(&nbd->config_lock);
1967         /*
1968          * Make sure recv thread has finished, so it does not drop the last
1969          * config ref and try to destroy the workqueue from inside the work
1970          * queue.
1971          */
1972         if (nbd->recv_workq)
1973                 flush_workqueue(nbd->recv_workq);
1974         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1975                                &nbd->config->runtime_flags))
1976                 nbd_config_put(nbd);
1977 }
1978
1979 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
1980 {
1981         struct nbd_device *nbd;
1982         int index;
1983
1984         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1985                 return -EPERM;
1986
1987         if (!info->attrs[NBD_ATTR_INDEX]) {
1988                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
1989                 return -EINVAL;
1990         }
1991         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1992         mutex_lock(&nbd_index_mutex);
1993         nbd = idr_find(&nbd_index_idr, index);
1994         if (!nbd) {
1995                 mutex_unlock(&nbd_index_mutex);
1996                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1997                        index);
1998                 return -EINVAL;
1999         }
2000         if (!refcount_inc_not_zero(&nbd->refs)) {
2001                 mutex_unlock(&nbd_index_mutex);
2002                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2003                        index);
2004                 return -EINVAL;
2005         }
2006         mutex_unlock(&nbd_index_mutex);
2007         if (!refcount_inc_not_zero(&nbd->config_refs))
2008                 goto put_nbd;
2009         nbd_disconnect_and_put(nbd);
2010         nbd_config_put(nbd);
2011 put_nbd:
2012         nbd_put(nbd);
2013         return 0;
2014 }
2015
2016 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2017 {
2018         struct nbd_device *nbd = NULL;
2019         struct nbd_config *config;
2020         int index;
2021         int ret = 0;
2022         bool put_dev = false;
2023
2024         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2025                 return -EPERM;
2026
2027         if (!info->attrs[NBD_ATTR_INDEX]) {
2028                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2029                 return -EINVAL;
2030         }
2031         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2032         mutex_lock(&nbd_index_mutex);
2033         nbd = idr_find(&nbd_index_idr, index);
2034         if (!nbd) {
2035                 mutex_unlock(&nbd_index_mutex);
2036                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2037                        index);
2038                 return -EINVAL;
2039         }
2040         if (!refcount_inc_not_zero(&nbd->refs)) {
2041                 mutex_unlock(&nbd_index_mutex);
2042                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2043                        index);
2044                 return -EINVAL;
2045         }
2046         mutex_unlock(&nbd_index_mutex);
2047
2048         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2049                 dev_err(nbd_to_dev(nbd),
2050                         "not configured, cannot reconfigure\n");
2051                 nbd_put(nbd);
2052                 return -EINVAL;
2053         }
2054
2055         mutex_lock(&nbd->config_lock);
2056         config = nbd->config;
2057         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2058             !nbd->task_recv) {
2059                 dev_err(nbd_to_dev(nbd),
2060                         "not configured, cannot reconfigure\n");
2061                 ret = -EINVAL;
2062                 goto out;
2063         }
2064
2065         ret = nbd_genl_size_set(info, nbd);
2066         if (ret)
2067                 goto out;
2068
2069         if (info->attrs[NBD_ATTR_TIMEOUT])
2070                 nbd_set_cmd_timeout(nbd,
2071                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2072         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2073                 config->dead_conn_timeout =
2074                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2075                 config->dead_conn_timeout *= HZ;
2076         }
2077         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2078                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2079                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2080                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2081                                               &nbd->flags))
2082                                 put_dev = true;
2083                 } else {
2084                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2085                                                &nbd->flags))
2086                                 refcount_inc(&nbd->refs);
2087                 }
2088
2089                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2090                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2091                                         &config->runtime_flags);
2092                 } else {
2093                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2094                                         &config->runtime_flags);
2095                 }
2096         }
2097
2098         if (info->attrs[NBD_ATTR_SOCKETS]) {
2099                 struct nlattr *attr;
2100                 int rem, fd;
2101
2102                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2103                                     rem) {
2104                         struct nlattr *socks[NBD_SOCK_MAX+1];
2105
2106                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2107                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2108                                 ret = -EINVAL;
2109                                 goto out;
2110                         }
2111                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2112                                                           attr,
2113                                                           nbd_sock_policy,
2114                                                           info->extack);
2115                         if (ret != 0) {
2116                                 printk(KERN_ERR "nbd: error processing sock list\n");
2117                                 ret = -EINVAL;
2118                                 goto out;
2119                         }
2120                         if (!socks[NBD_SOCK_FD])
2121                                 continue;
2122                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2123                         ret = nbd_reconnect_socket(nbd, fd);
2124                         if (ret) {
2125                                 if (ret == -ENOSPC)
2126                                         ret = 0;
2127                                 goto out;
2128                         }
2129                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2130                 }
2131         }
2132 out:
2133         mutex_unlock(&nbd->config_lock);
2134         nbd_config_put(nbd);
2135         nbd_put(nbd);
2136         if (put_dev)
2137                 nbd_put(nbd);
2138         return ret;
2139 }
2140
2141 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2142         {
2143                 .cmd    = NBD_CMD_CONNECT,
2144                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2145                 .doit   = nbd_genl_connect,
2146         },
2147         {
2148                 .cmd    = NBD_CMD_DISCONNECT,
2149                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2150                 .doit   = nbd_genl_disconnect,
2151         },
2152         {
2153                 .cmd    = NBD_CMD_RECONFIGURE,
2154                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2155                 .doit   = nbd_genl_reconfigure,
2156         },
2157         {
2158                 .cmd    = NBD_CMD_STATUS,
2159                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2160                 .doit   = nbd_genl_status,
2161         },
2162 };
2163
2164 static const struct genl_multicast_group nbd_mcast_grps[] = {
2165         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2166 };
2167
2168 static struct genl_family nbd_genl_family __ro_after_init = {
2169         .hdrsize        = 0,
2170         .name           = NBD_GENL_FAMILY_NAME,
2171         .version        = NBD_GENL_VERSION,
2172         .module         = THIS_MODULE,
2173         .small_ops      = nbd_connect_genl_ops,
2174         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2175         .maxattr        = NBD_ATTR_MAX,
2176         .policy = nbd_attr_policy,
2177         .mcgrps         = nbd_mcast_grps,
2178         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2179 };
2180
2181 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2182 {
2183         struct nlattr *dev_opt;
2184         u8 connected = 0;
2185         int ret;
2186
2187         /* This is a little racey, but for status it's ok.  The
2188          * reason we don't take a ref here is because we can't
2189          * take a ref in the index == -1 case as we would need
2190          * to put under the nbd_index_mutex, which could
2191          * deadlock if we are configured to remove ourselves
2192          * once we're disconnected.
2193          */
2194         if (refcount_read(&nbd->config_refs))
2195                 connected = 1;
2196         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2197         if (!dev_opt)
2198                 return -EMSGSIZE;
2199         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2200         if (ret)
2201                 return -EMSGSIZE;
2202         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2203                          connected);
2204         if (ret)
2205                 return -EMSGSIZE;
2206         nla_nest_end(reply, dev_opt);
2207         return 0;
2208 }
2209
2210 static int status_cb(int id, void *ptr, void *data)
2211 {
2212         struct nbd_device *nbd = ptr;
2213         return populate_nbd_status(nbd, (struct sk_buff *)data);
2214 }
2215
2216 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2217 {
2218         struct nlattr *dev_list;
2219         struct sk_buff *reply;
2220         void *reply_head;
2221         size_t msg_size;
2222         int index = -1;
2223         int ret = -ENOMEM;
2224
2225         if (info->attrs[NBD_ATTR_INDEX])
2226                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2227
2228         mutex_lock(&nbd_index_mutex);
2229
2230         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2231                                   nla_attr_size(sizeof(u8)));
2232         msg_size *= (index == -1) ? nbd_total_devices : 1;
2233
2234         reply = genlmsg_new(msg_size, GFP_KERNEL);
2235         if (!reply)
2236                 goto out;
2237         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2238                                        NBD_CMD_STATUS);
2239         if (!reply_head) {
2240                 nlmsg_free(reply);
2241                 goto out;
2242         }
2243
2244         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2245         if (index == -1) {
2246                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2247                 if (ret) {
2248                         nlmsg_free(reply);
2249                         goto out;
2250                 }
2251         } else {
2252                 struct nbd_device *nbd;
2253                 nbd = idr_find(&nbd_index_idr, index);
2254                 if (nbd) {
2255                         ret = populate_nbd_status(nbd, reply);
2256                         if (ret) {
2257                                 nlmsg_free(reply);
2258                                 goto out;
2259                         }
2260                 }
2261         }
2262         nla_nest_end(reply, dev_list);
2263         genlmsg_end(reply, reply_head);
2264         ret = genlmsg_reply(reply, info);
2265 out:
2266         mutex_unlock(&nbd_index_mutex);
2267         return ret;
2268 }
2269
2270 static void nbd_connect_reply(struct genl_info *info, int index)
2271 {
2272         struct sk_buff *skb;
2273         void *msg_head;
2274         int ret;
2275
2276         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2277         if (!skb)
2278                 return;
2279         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2280                                      NBD_CMD_CONNECT);
2281         if (!msg_head) {
2282                 nlmsg_free(skb);
2283                 return;
2284         }
2285         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2286         if (ret) {
2287                 nlmsg_free(skb);
2288                 return;
2289         }
2290         genlmsg_end(skb, msg_head);
2291         genlmsg_reply(skb, info);
2292 }
2293
2294 static void nbd_mcast_index(int index)
2295 {
2296         struct sk_buff *skb;
2297         void *msg_head;
2298         int ret;
2299
2300         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2301         if (!skb)
2302                 return;
2303         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2304                                      NBD_CMD_LINK_DEAD);
2305         if (!msg_head) {
2306                 nlmsg_free(skb);
2307                 return;
2308         }
2309         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2310         if (ret) {
2311                 nlmsg_free(skb);
2312                 return;
2313         }
2314         genlmsg_end(skb, msg_head);
2315         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2316 }
2317
2318 static void nbd_dead_link_work(struct work_struct *work)
2319 {
2320         struct link_dead_args *args = container_of(work, struct link_dead_args,
2321                                                    work);
2322         nbd_mcast_index(args->index);
2323         kfree(args);
2324 }
2325
2326 static int __init nbd_init(void)
2327 {
2328         int i;
2329
2330         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2331
2332         if (max_part < 0) {
2333                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2334                 return -EINVAL;
2335         }
2336
2337         part_shift = 0;
2338         if (max_part > 0) {
2339                 part_shift = fls(max_part);
2340
2341                 /*
2342                  * Adjust max_part according to part_shift as it is exported
2343                  * to user space so that user can know the max number of
2344                  * partition kernel should be able to manage.
2345                  *
2346                  * Note that -1 is required because partition 0 is reserved
2347                  * for the whole disk.
2348                  */
2349                 max_part = (1UL << part_shift) - 1;
2350         }
2351
2352         if ((1UL << part_shift) > DISK_MAX_PARTS)
2353                 return -EINVAL;
2354
2355         if (nbds_max > 1UL << (MINORBITS - part_shift))
2356                 return -EINVAL;
2357
2358         if (register_blkdev(NBD_MAJOR, "nbd"))
2359                 return -EIO;
2360
2361         if (genl_register_family(&nbd_genl_family)) {
2362                 unregister_blkdev(NBD_MAJOR, "nbd");
2363                 return -EINVAL;
2364         }
2365         nbd_dbg_init();
2366
2367         mutex_lock(&nbd_index_mutex);
2368         for (i = 0; i < nbds_max; i++)
2369                 nbd_dev_add(i);
2370         mutex_unlock(&nbd_index_mutex);
2371         return 0;
2372 }
2373
2374 static int nbd_exit_cb(int id, void *ptr, void *data)
2375 {
2376         struct list_head *list = (struct list_head *)data;
2377         struct nbd_device *nbd = ptr;
2378
2379         list_add_tail(&nbd->list, list);
2380         return 0;
2381 }
2382
2383 static void __exit nbd_cleanup(void)
2384 {
2385         struct nbd_device *nbd;
2386         LIST_HEAD(del_list);
2387
2388         nbd_dbg_close();
2389
2390         mutex_lock(&nbd_index_mutex);
2391         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2392         mutex_unlock(&nbd_index_mutex);
2393
2394         while (!list_empty(&del_list)) {
2395                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2396                 list_del_init(&nbd->list);
2397                 if (refcount_read(&nbd->refs) != 1)
2398                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2399                 nbd_put(nbd);
2400         }
2401
2402         idr_destroy(&nbd_index_idr);
2403         genl_unregister_family(&nbd_genl_family);
2404         unregister_blkdev(NBD_MAJOR, "nbd");
2405 }
2406
2407 module_init(nbd_init);
2408 module_exit(nbd_cleanup);
2409
2410 MODULE_DESCRIPTION("Network Block Device");
2411 MODULE_LICENSE("GPL");
2412
2413 module_param(nbds_max, int, 0444);
2414 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2415 module_param(max_part, int, 0444);
2416 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");