drivers/block/nbd.c

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