drivers/block/nbd.c

   1 /*
   2  * Network block device - make block devices work over TCP
   3  *
   4  * Note that you can not swap over this thing, yet. Seems to work but
   5  * deadlocks sometimes - you can not swap over TCP in general.
   6  *
   7  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   8  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
   9  *
  10  * This file is released under GPLv2 or later.
  11  *
  12  * (part of code stolen from loop.c)
  13  */
  14
  15 #include <linux/major.h>
  16
  17 #include <linux/blkdev.h>
  18 #include <linux/module.h>
  19 #include <linux/init.h>
  20 #include <linux/sched.h>
  21 #include <linux/sched/mm.h>
  22 #include <linux/fs.h>
  23 #include <linux/bio.h>
  24 #include <linux/stat.h>
  25 #include <linux/errno.h>
  26 #include <linux/file.h>
  27 #include <linux/ioctl.h>
  28 #include <linux/mutex.h>
  29 #include <linux/compiler.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 static DEFINE_IDR(nbd_index_idr);
  48 static DEFINE_MUTEX(nbd_index_mutex);
  49 static int nbd_total_devices = 0;
  50
  51 struct nbd_sock {
  52         struct socket *sock;
  53         struct mutex tx_lock;
  54         struct request *pending;
  55         int sent;
  56         bool dead;
  57         int fallback_index;
  58         int cookie;
  59 };
  60
  61 struct recv_thread_args {
  62         struct work_struct work;
  63         struct nbd_device *nbd;
  64         int index;
  65 };
  66
  67 struct link_dead_args {
  68         struct work_struct work;
  69         int index;
  70 };
  71
  72 #define NBD_TIMEDOUT                    0
  73 #define NBD_DISCONNECT_REQUESTED        1
  74 #define NBD_DISCONNECTED                2
  75 #define NBD_HAS_PID_FILE                3
  76 #define NBD_HAS_CONFIG_REF              4
  77 #define NBD_BOUND                       5
  78 #define NBD_DESTROY_ON_DISCONNECT       6
  79 #define NBD_DISCONNECT_ON_CLOSE         7
  80
  81 struct nbd_config {
  82         u32 flags;
  83         unsigned long runtime_flags;
  84         u64 dead_conn_timeout;
  85
  86         struct nbd_sock **socks;
  87         int num_connections;
  88         atomic_t live_connections;
  89         wait_queue_head_t conn_wait;
  90
  91         atomic_t recv_threads;
  92         wait_queue_head_t recv_wq;
  93         loff_t blksize;
  94         loff_t bytesize;
  95 #if IS_ENABLED(CONFIG_DEBUG_FS)
  96         struct dentry *dbg_dir;
  97 #endif
  98 };
  99
 100 struct nbd_device {
 101         struct blk_mq_tag_set tag_set;
 102
 103         int index;
 104         refcount_t config_refs;
 105         refcount_t refs;
 106         struct nbd_config *config;
 107         struct mutex config_lock;
 108         struct gendisk *disk;
 109
 110         struct list_head list;
 111         struct task_struct *task_recv;
 112         struct task_struct *task_setup;
 113 };
 114
 115 #define NBD_CMD_REQUEUED        1
 116
 117 struct nbd_cmd {
 118         struct nbd_device *nbd;
 119         struct mutex lock;
 120         int index;
 121         int cookie;
 122         blk_status_t status;
 123         unsigned long flags;
 124         u32 cmd_cookie;
 125 };
 126
 127 #if IS_ENABLED(CONFIG_DEBUG_FS)
 128 static struct dentry *nbd_dbg_dir;
 129 #endif
 130
 131 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 132
 133 #define NBD_MAGIC 0x68797548
 134
 135 static unsigned int nbds_max = 16;
 136 static int max_part = 16;
 137 static struct workqueue_struct *recv_workqueue;
 138 static int part_shift;
 139
 140 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 141 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 142 static void nbd_config_put(struct nbd_device *nbd);
 143 static void nbd_connect_reply(struct genl_info *info, int index);
 144 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 145 static void nbd_dead_link_work(struct work_struct *work);
 146 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 147
 148 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 149 {
 150         return disk_to_dev(nbd->disk);
 151 }
 152
 153 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 154 {
 155         struct request *req = blk_mq_rq_from_pdu(cmd);
 156
 157         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 158                 blk_mq_requeue_request(req, true);
 159 }
 160
 161 #define NBD_COOKIE_BITS 32
 162
 163 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 164 {
 165         struct request *req = blk_mq_rq_from_pdu(cmd);
 166         u32 tag = blk_mq_unique_tag(req);
 167         u64 cookie = cmd->cmd_cookie;
 168
 169         return (cookie << NBD_COOKIE_BITS) | tag;
 170 }
 171
 172 static u32 nbd_handle_to_tag(u64 handle)
 173 {
 174         return (u32)handle;
 175 }
 176
 177 static u32 nbd_handle_to_cookie(u64 handle)
 178 {
 179         return (u32)(handle >> NBD_COOKIE_BITS);
 180 }
 181
 182 static const char *nbdcmd_to_ascii(int cmd)
 183 {
 184         switch (cmd) {
 185         case  NBD_CMD_READ: return "read";
 186         case NBD_CMD_WRITE: return "write";
 187         case  NBD_CMD_DISC: return "disconnect";
 188         case NBD_CMD_FLUSH: return "flush";
 189         case  NBD_CMD_TRIM: return "trim/discard";
 190         }
 191         return "invalid";
 192 }
 193
 194 static ssize_t pid_show(struct device *dev,
 195                         struct device_attribute *attr, char *buf)
 196 {
 197         struct gendisk *disk = dev_to_disk(dev);
 198         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 199
 200         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 201 }
 202
 203 static const struct device_attribute pid_attr = {
 204         .attr = { .name = "pid", .mode = 0444},
 205         .show = pid_show,
 206 };
 207
 208 static void nbd_dev_remove(struct nbd_device *nbd)
 209 {
 210         struct gendisk *disk = nbd->disk;
 211         struct request_queue *q;
 212
 213         if (disk) {
 214                 q = disk->queue;
 215                 del_gendisk(disk);
 216                 blk_cleanup_queue(q);
 217                 blk_mq_free_tag_set(&nbd->tag_set);
 218                 disk->private_data = NULL;
 219                 put_disk(disk);
 220         }
 221         kfree(nbd);
 222 }
 223
 224 static void nbd_put(struct nbd_device *nbd)
 225 {
 226         if (refcount_dec_and_mutex_lock(&nbd->refs,
 227                                         &nbd_index_mutex)) {
 228                 idr_remove(&nbd_index_idr, nbd->index);
 229                 mutex_unlock(&nbd_index_mutex);
 230                 nbd_dev_remove(nbd);
 231         }
 232 }
 233
 234 static int nbd_disconnected(struct nbd_config *config)
 235 {
 236         return test_bit(NBD_DISCONNECTED, &config->runtime_flags) ||
 237                 test_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags);
 238 }
 239
 240 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 241                                 int notify)
 242 {
 243         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 244                 struct link_dead_args *args;
 245                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 246                 if (args) {
 247                         INIT_WORK(&args->work, nbd_dead_link_work);
 248                         args->index = nbd->index;
 249                         queue_work(system_wq, &args->work);
 250                 }
 251         }
 252         if (!nsock->dead) {
 253                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 254                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 255                         if (test_and_clear_bit(NBD_DISCONNECT_REQUESTED,
 256                                                &nbd->config->runtime_flags)) {
 257                                 set_bit(NBD_DISCONNECTED,
 258                                         &nbd->config->runtime_flags);
 259                                 dev_info(nbd_to_dev(nbd),
 260                                         "Disconnected due to user request.\n");
 261                         }
 262                 }
 263         }
 264         nsock->dead = true;
 265         nsock->pending = NULL;
 266         nsock->sent = 0;
 267 }
 268
 269 static void nbd_size_clear(struct nbd_device *nbd)
 270 {
 271         if (nbd->config->bytesize) {
 272                 set_capacity(nbd->disk, 0);
 273                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 274         }
 275 }
 276
 277 static void nbd_size_update(struct nbd_device *nbd)
 278 {
 279         struct nbd_config *config = nbd->config;
 280         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 281
 282         if (config->flags & NBD_FLAG_SEND_TRIM) {
 283                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 284                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 285                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 286         }
 287         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 288         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 289         set_capacity(nbd->disk, config->bytesize >> 9);
 290         if (bdev) {
 291                 if (bdev->bd_disk)
 292                         bd_set_size(bdev, config->bytesize);
 293                 else
 294                         bdev->bd_invalidated = 1;
 295                 bdput(bdev);
 296         }
 297         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 298 }
 299
 300 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 301                          loff_t nr_blocks)
 302 {
 303         struct nbd_config *config = nbd->config;
 304         config->blksize = blocksize;
 305         config->bytesize = blocksize * nr_blocks;
 306         if (nbd->task_recv != NULL)
 307                 nbd_size_update(nbd);
 308 }
 309
 310 static void nbd_complete_rq(struct request *req)
 311 {
 312         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 313
 314         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 315                 cmd->status ? "failed" : "done");
 316
 317         blk_mq_end_request(req, cmd->status);
 318 }
 319
 320 /*
 321  * Forcibly shutdown the socket causing all listeners to error
 322  */
 323 static void sock_shutdown(struct nbd_device *nbd)
 324 {
 325         struct nbd_config *config = nbd->config;
 326         int i;
 327
 328         if (config->num_connections == 0)
 329                 return;
 330         if (test_and_set_bit(NBD_DISCONNECTED, &config->runtime_flags))
 331                 return;
 332
 333         for (i = 0; i < config->num_connections; i++) {
 334                 struct nbd_sock *nsock = config->socks[i];
 335                 mutex_lock(&nsock->tx_lock);
 336                 nbd_mark_nsock_dead(nbd, nsock, 0);
 337                 mutex_unlock(&nsock->tx_lock);
 338         }
 339         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 340 }
 341
 342 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 343                                                  bool reserved)
 344 {
 345         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 346         struct nbd_device *nbd = cmd->nbd;
 347         struct nbd_config *config;
 348
 349         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 350                 cmd->status = BLK_STS_TIMEOUT;
 351                 goto done;
 352         }
 353         config = nbd->config;
 354
 355         if (!mutex_trylock(&cmd->lock))
 356                 return BLK_EH_RESET_TIMER;
 357
 358         if (config->num_connections > 1) {
 359                 dev_err_ratelimited(nbd_to_dev(nbd),
 360                                     "Connection timed out, retrying (%d/%d alive)\n",
 361                                     atomic_read(&config->live_connections),
 362                                     config->num_connections);
 363                 /*
 364                  * Hooray we have more connections, requeue this IO, the submit
 365                  * path will put it on a real connection.
 366                  */
 367                 if (config->socks && config->num_connections > 1) {
 368                         if (cmd->index < config->num_connections) {
 369                                 struct nbd_sock *nsock =
 370                                         config->socks[cmd->index];
 371                                 mutex_lock(&nsock->tx_lock);
 372                                 /* We can have multiple outstanding requests, so
 373                                  * we don't want to mark the nsock dead if we've
 374                                  * already reconnected with a new socket, so
 375                                  * only mark it dead if its the same socket we
 376                                  * were sent out on.
 377                                  */
 378                                 if (cmd->cookie == nsock->cookie)
 379                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 380                                 mutex_unlock(&nsock->tx_lock);
 381                         }
 382                         mutex_unlock(&cmd->lock);
 383                         nbd_requeue_cmd(cmd);
 384                         nbd_config_put(nbd);
 385                         return BLK_EH_DONE;
 386                 }
 387         } else {
 388                 dev_err_ratelimited(nbd_to_dev(nbd),
 389                                     "Connection timed out\n");
 390         }
 391         set_bit(NBD_TIMEDOUT, &config->runtime_flags);
 392         cmd->status = BLK_STS_IOERR;
 393         mutex_unlock(&cmd->lock);
 394         sock_shutdown(nbd);
 395         nbd_config_put(nbd);
 396 done:
 397         blk_mq_complete_request(req);
 398         return BLK_EH_DONE;
 399 }
 400
 401 /*
 402  *  Send or receive packet.
 403  */
 404 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 405                      struct iov_iter *iter, int msg_flags, int *sent)
 406 {
 407         struct nbd_config *config = nbd->config;
 408         struct socket *sock = config->socks[index]->sock;
 409         int result;
 410         struct msghdr msg;
 411         unsigned int noreclaim_flag;
 412
 413         if (unlikely(!sock)) {
 414                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 415                         "Attempted %s on closed socket in sock_xmit\n",
 416                         (send ? "send" : "recv"));
 417                 return -EINVAL;
 418         }
 419
 420         msg.msg_iter = *iter;
 421
 422         noreclaim_flag = memalloc_noreclaim_save();
 423         do {
 424                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 425                 msg.msg_name = NULL;
 426                 msg.msg_namelen = 0;
 427                 msg.msg_control = NULL;
 428                 msg.msg_controllen = 0;
 429                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 430
 431                 if (send)
 432                         result = sock_sendmsg(sock, &msg);
 433                 else
 434                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 435
 436                 if (result <= 0) {
 437                         if (result == 0)
 438                                 result = -EPIPE; /* short read */
 439                         break;
 440                 }
 441                 if (sent)
 442                         *sent += result;
 443         } while (msg_data_left(&msg));
 444
 445         memalloc_noreclaim_restore(noreclaim_flag);
 446
 447         return result;
 448 }
 449
 450 /*
 451  * Different settings for sk->sk_sndtimeo can result in different return values
 452  * if there is a signal pending when we enter sendmsg, because reasons?
 453  */
 454 static inline int was_interrupted(int result)
 455 {
 456         return result == -ERESTARTSYS || result == -EINTR;
 457 }
 458
 459 /* always call with the tx_lock held */
 460 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 461 {
 462         struct request *req = blk_mq_rq_from_pdu(cmd);
 463         struct nbd_config *config = nbd->config;
 464         struct nbd_sock *nsock = config->socks[index];
 465         int result;
 466         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 467         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 468         struct iov_iter from;
 469         unsigned long size = blk_rq_bytes(req);
 470         struct bio *bio;
 471         u64 handle;
 472         u32 type;
 473         u32 nbd_cmd_flags = 0;
 474         int sent = nsock->sent, skip = 0;
 475
 476         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 477
 478         switch (req_op(req)) {
 479         case REQ_OP_DISCARD:
 480                 type = NBD_CMD_TRIM;
 481                 break;
 482         case REQ_OP_FLUSH:
 483                 type = NBD_CMD_FLUSH;
 484                 break;
 485         case REQ_OP_WRITE:
 486                 type = NBD_CMD_WRITE;
 487                 break;
 488         case REQ_OP_READ:
 489                 type = NBD_CMD_READ;
 490                 break;
 491         default:
 492                 return -EIO;
 493         }
 494
 495         if (rq_data_dir(req) == WRITE &&
 496             (config->flags & NBD_FLAG_READ_ONLY)) {
 497                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 498                                     "Write on read-only\n");
 499                 return -EIO;
 500         }
 501
 502         if (req->cmd_flags & REQ_FUA)
 503                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 504
 505         /* We did a partial send previously, and we at least sent the whole
 506          * request struct, so just go and send the rest of the pages in the
 507          * request.
 508          */
 509         if (sent) {
 510                 if (sent >= sizeof(request)) {
 511                         skip = sent - sizeof(request);
 512                         goto send_pages;
 513                 }
 514                 iov_iter_advance(&from, sent);
 515         } else {
 516                 cmd->cmd_cookie++;
 517         }
 518         cmd->index = index;
 519         cmd->cookie = nsock->cookie;
 520         request.type = htonl(type | nbd_cmd_flags);
 521         if (type != NBD_CMD_FLUSH) {
 522                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 523                 request.len = htonl(size);
 524         }
 525         handle = nbd_cmd_handle(cmd);
 526         memcpy(request.handle, &handle, sizeof(handle));
 527
 528         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 529                 req, nbdcmd_to_ascii(type),
 530                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 531         result = sock_xmit(nbd, index, 1, &from,
 532                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 533         if (result <= 0) {
 534                 if (was_interrupted(result)) {
 535                         /* If we havne't sent anything we can just return BUSY,
 536                          * however if we have sent something we need to make
 537                          * sure we only allow this req to be sent until we are
 538                          * completely done.
 539                          */
 540                         if (sent) {
 541                                 nsock->pending = req;
 542                                 nsock->sent = sent;
 543                         }
 544                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 545                         return BLK_STS_RESOURCE;
 546                 }
 547                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 548                         "Send control failed (result %d)\n", result);
 549                 return -EAGAIN;
 550         }
 551 send_pages:
 552         if (type != NBD_CMD_WRITE)
 553                 goto out;
 554
 555         bio = req->bio;
 556         while (bio) {
 557                 struct bio *next = bio->bi_next;
 558                 struct bvec_iter iter;
 559                 struct bio_vec bvec;
 560
 561                 bio_for_each_segment(bvec, bio, iter) {
 562                         bool is_last = !next && bio_iter_last(bvec, iter);
 563                         int flags = is_last ? 0 : MSG_MORE;
 564
 565                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 566                                 req, bvec.bv_len);
 567                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 568                         if (skip) {
 569                                 if (skip >= iov_iter_count(&from)) {
 570                                         skip -= iov_iter_count(&from);
 571                                         continue;
 572                                 }
 573                                 iov_iter_advance(&from, skip);
 574                                 skip = 0;
 575                         }
 576                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 577                         if (result <= 0) {
 578                                 if (was_interrupted(result)) {
 579                                         /* We've already sent the header, we
 580                                          * have no choice but to set pending and
 581                                          * return BUSY.
 582                                          */
 583                                         nsock->pending = req;
 584                                         nsock->sent = sent;
 585                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 586                                         return BLK_STS_RESOURCE;
 587                                 }
 588                                 dev_err(disk_to_dev(nbd->disk),
 589                                         "Send data failed (result %d)\n",
 590                                         result);
 591                                 return -EAGAIN;
 592                         }
 593                         /*
 594                          * The completion might already have come in,
 595                          * so break for the last one instead of letting
 596                          * the iterator do it. This prevents use-after-free
 597                          * of the bio.
 598                          */
 599                         if (is_last)
 600                                 break;
 601                 }
 602                 bio = next;
 603         }
 604 out:
 605         nsock->pending = NULL;
 606         nsock->sent = 0;
 607         return 0;
 608 }
 609
 610 /* NULL returned = something went wrong, inform userspace */
 611 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 612 {
 613         struct nbd_config *config = nbd->config;
 614         int result;
 615         struct nbd_reply reply;
 616         struct nbd_cmd *cmd;
 617         struct request *req = NULL;
 618         u64 handle;
 619         u16 hwq;
 620         u32 tag;
 621         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 622         struct iov_iter to;
 623         int ret = 0;
 624
 625         reply.magic = 0;
 626         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 627         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 628         if (result <= 0) {
 629                 if (!nbd_disconnected(config))
 630                         dev_err(disk_to_dev(nbd->disk),
 631                                 "Receive control failed (result %d)\n", result);
 632                 return ERR_PTR(result);
 633         }
 634
 635         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 636                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 637                                 (unsigned long)ntohl(reply.magic));
 638                 return ERR_PTR(-EPROTO);
 639         }
 640
 641         memcpy(&handle, reply.handle, sizeof(handle));
 642         tag = nbd_handle_to_tag(handle);
 643         hwq = blk_mq_unique_tag_to_hwq(tag);
 644         if (hwq < nbd->tag_set.nr_hw_queues)
 645                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 646                                        blk_mq_unique_tag_to_tag(tag));
 647         if (!req || !blk_mq_request_started(req)) {
 648                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 649                         tag, req);
 650                 return ERR_PTR(-ENOENT);
 651         }
 652         cmd = blk_mq_rq_to_pdu(req);
 653
 654         mutex_lock(&cmd->lock);
 655         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 656                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 657                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 658                 ret = -ENOENT;
 659                 goto out;
 660         }
 661         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 662                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 663                         req);
 664                 ret = -ENOENT;
 665                 goto out;
 666         }
 667         if (ntohl(reply.error)) {
 668                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 669                         ntohl(reply.error));
 670                 cmd->status = BLK_STS_IOERR;
 671                 goto out;
 672         }
 673
 674         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 675         if (rq_data_dir(req) != WRITE) {
 676                 struct req_iterator iter;
 677                 struct bio_vec bvec;
 678
 679                 rq_for_each_segment(bvec, req, iter) {
 680                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 681                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 682                         if (result <= 0) {
 683                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 684                                         result);
 685                                 /*
 686                                  * If we've disconnected or we only have 1
 687                                  * connection then we need to make sure we
 688                                  * complete this request, otherwise error out
 689                                  * and let the timeout stuff handle resubmitting
 690                                  * this request onto another connection.
 691                                  */
 692                                 if (nbd_disconnected(config) ||
 693                                     config->num_connections <= 1) {
 694                                         cmd->status = BLK_STS_IOERR;
 695                                         goto out;
 696                                 }
 697                                 ret = -EIO;
 698                                 goto out;
 699                         }
 700                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 701                                 req, bvec.bv_len);
 702                 }
 703         }
 704 out:
 705         mutex_unlock(&cmd->lock);
 706         return ret ? ERR_PTR(ret) : cmd;
 707 }
 708
 709 static void recv_work(struct work_struct *work)
 710 {
 711         struct recv_thread_args *args = container_of(work,
 712                                                      struct recv_thread_args,
 713                                                      work);
 714         struct nbd_device *nbd = args->nbd;
 715         struct nbd_config *config = nbd->config;
 716         struct nbd_cmd *cmd;
 717
 718         while (1) {
 719                 cmd = nbd_read_stat(nbd, args->index);
 720                 if (IS_ERR(cmd)) {
 721                         struct nbd_sock *nsock = config->socks[args->index];
 722
 723                         mutex_lock(&nsock->tx_lock);
 724                         nbd_mark_nsock_dead(nbd, nsock, 1);
 725                         mutex_unlock(&nsock->tx_lock);
 726                         break;
 727                 }
 728
 729                 blk_mq_complete_request(blk_mq_rq_from_pdu(cmd));
 730         }
 731         atomic_dec(&config->recv_threads);
 732         wake_up(&config->recv_wq);
 733         nbd_config_put(nbd);
 734         kfree(args);
 735 }
 736
 737 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 738 {
 739         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 740
 741         cmd->status = BLK_STS_IOERR;
 742         blk_mq_complete_request(req);
 743         return true;
 744 }
 745
 746 static void nbd_clear_que(struct nbd_device *nbd)
 747 {
 748         blk_mq_quiesce_queue(nbd->disk->queue);
 749         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 750         blk_mq_unquiesce_queue(nbd->disk->queue);
 751         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 752 }
 753
 754 static int find_fallback(struct nbd_device *nbd, int index)
 755 {
 756         struct nbd_config *config = nbd->config;
 757         int new_index = -1;
 758         struct nbd_sock *nsock = config->socks[index];
 759         int fallback = nsock->fallback_index;
 760
 761         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 762                 return new_index;
 763
 764         if (config->num_connections <= 1) {
 765                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 766                                     "Attempted send on invalid socket\n");
 767                 return new_index;
 768         }
 769
 770         if (fallback >= 0 && fallback < config->num_connections &&
 771             !config->socks[fallback]->dead)
 772                 return fallback;
 773
 774         if (nsock->fallback_index < 0 ||
 775             nsock->fallback_index >= config->num_connections ||
 776             config->socks[nsock->fallback_index]->dead) {
 777                 int i;
 778                 for (i = 0; i < config->num_connections; i++) {
 779                         if (i == index)
 780                                 continue;
 781                         if (!config->socks[i]->dead) {
 782                                 new_index = i;
 783                                 break;
 784                         }
 785                 }
 786                 nsock->fallback_index = new_index;
 787                 if (new_index < 0) {
 788                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 789                                             "Dead connection, failed to find a fallback\n");
 790                         return new_index;
 791                 }
 792         }
 793         new_index = nsock->fallback_index;
 794         return new_index;
 795 }
 796
 797 static int wait_for_reconnect(struct nbd_device *nbd)
 798 {
 799         struct nbd_config *config = nbd->config;
 800         if (!config->dead_conn_timeout)
 801                 return 0;
 802         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 803                 return 0;
 804         return wait_event_timeout(config->conn_wait,
 805                                   atomic_read(&config->live_connections) > 0,
 806                                   config->dead_conn_timeout) > 0;
 807 }
 808
 809 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 810 {
 811         struct request *req = blk_mq_rq_from_pdu(cmd);
 812         struct nbd_device *nbd = cmd->nbd;
 813         struct nbd_config *config;
 814         struct nbd_sock *nsock;
 815         int ret;
 816
 817         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 818                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 819                                     "Socks array is empty\n");
 820                 blk_mq_start_request(req);
 821                 return -EINVAL;
 822         }
 823         config = nbd->config;
 824
 825         if (index >= config->num_connections) {
 826                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 827                                     "Attempted send on invalid socket\n");
 828                 nbd_config_put(nbd);
 829                 blk_mq_start_request(req);
 830                 return -EINVAL;
 831         }
 832         cmd->status = BLK_STS_OK;
 833 again:
 834         nsock = config->socks[index];
 835         mutex_lock(&nsock->tx_lock);
 836         if (nsock->dead) {
 837                 int old_index = index;
 838                 index = find_fallback(nbd, index);
 839                 mutex_unlock(&nsock->tx_lock);
 840                 if (index < 0) {
 841                         if (wait_for_reconnect(nbd)) {
 842                                 index = old_index;
 843                                 goto again;
 844                         }
 845                         /* All the sockets should already be down at this point,
 846                          * we just want to make sure that DISCONNECTED is set so
 847                          * any requests that come in that were queue'ed waiting
 848                          * for the reconnect timer don't trigger the timer again
 849                          * and instead just error out.
 850                          */
 851                         sock_shutdown(nbd);
 852                         nbd_config_put(nbd);
 853                         blk_mq_start_request(req);
 854                         return -EIO;
 855                 }
 856                 goto again;
 857         }
 858
 859         /* Handle the case that we have a pending request that was partially
 860          * transmitted that _has_ to be serviced first.  We need to call requeue
 861          * here so that it gets put _after_ the request that is already on the
 862          * dispatch list.
 863          */
 864         blk_mq_start_request(req);
 865         if (unlikely(nsock->pending && nsock->pending != req)) {
 866                 nbd_requeue_cmd(cmd);
 867                 ret = 0;
 868                 goto out;
 869         }
 870         /*
 871          * Some failures are related to the link going down, so anything that
 872          * returns EAGAIN can be retried on a different socket.
 873          */
 874         ret = nbd_send_cmd(nbd, cmd, index);
 875         if (ret == -EAGAIN) {
 876                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 877                                     "Request send failed, requeueing\n");
 878                 nbd_mark_nsock_dead(nbd, nsock, 1);
 879                 nbd_requeue_cmd(cmd);
 880                 ret = 0;
 881         }
 882 out:
 883         mutex_unlock(&nsock->tx_lock);
 884         nbd_config_put(nbd);
 885         return ret;
 886 }
 887
 888 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 889                         const struct blk_mq_queue_data *bd)
 890 {
 891         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 892         int ret;
 893
 894         /*
 895          * Since we look at the bio's to send the request over the network we
 896          * need to make sure the completion work doesn't mark this request done
 897          * before we are done doing our send.  This keeps us from dereferencing
 898          * freed data if we have particularly fast completions (ie we get the
 899          * completion before we exit sock_xmit on the last bvec) or in the case
 900          * that the server is misbehaving (or there was an error) before we're
 901          * done sending everything over the wire.
 902          */
 903         mutex_lock(&cmd->lock);
 904         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 905
 906         /* We can be called directly from the user space process, which means we
 907          * could possibly have signals pending so our sendmsg will fail.  In
 908          * this case we need to return that we are busy, otherwise error out as
 909          * appropriate.
 910          */
 911         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 912         if (ret < 0)
 913                 ret = BLK_STS_IOERR;
 914         else if (!ret)
 915                 ret = BLK_STS_OK;
 916         mutex_unlock(&cmd->lock);
 917
 918         return ret;
 919 }
 920
 921 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
 922                           bool netlink)
 923 {
 924         struct nbd_config *config = nbd->config;
 925         struct socket *sock;
 926         struct nbd_sock **socks;
 927         struct nbd_sock *nsock;
 928         int err;
 929
 930         sock = sockfd_lookup(arg, &err);
 931         if (!sock)
 932                 return err;
 933
 934         if (!netlink && !nbd->task_setup &&
 935             !test_bit(NBD_BOUND, &config->runtime_flags))
 936                 nbd->task_setup = current;
 937
 938         if (!netlink &&
 939             (nbd->task_setup != current ||
 940              test_bit(NBD_BOUND, &config->runtime_flags))) {
 941                 dev_err(disk_to_dev(nbd->disk),
 942                         "Device being setup by another task");
 943                 sockfd_put(sock);
 944                 return -EBUSY;
 945         }
 946
 947         socks = krealloc(config->socks, (config->num_connections + 1) *
 948                          sizeof(struct nbd_sock *), GFP_KERNEL);
 949         if (!socks) {
 950                 sockfd_put(sock);
 951                 return -ENOMEM;
 952         }
 953         nsock = kzalloc(sizeof(struct nbd_sock), GFP_KERNEL);
 954         if (!nsock) {
 955                 sockfd_put(sock);
 956                 return -ENOMEM;
 957         }
 958
 959         config->socks = socks;
 960
 961         nsock->fallback_index = -1;
 962         nsock->dead = false;
 963         mutex_init(&nsock->tx_lock);
 964         nsock->sock = sock;
 965         nsock->pending = NULL;
 966         nsock->sent = 0;
 967         nsock->cookie = 0;
 968         socks[config->num_connections++] = nsock;
 969         atomic_inc(&config->live_connections);
 970
 971         return 0;
 972 }
 973
 974 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
 975 {
 976         struct nbd_config *config = nbd->config;
 977         struct socket *sock, *old;
 978         struct recv_thread_args *args;
 979         int i;
 980         int err;
 981
 982         sock = sockfd_lookup(arg, &err);
 983         if (!sock)
 984                 return err;
 985
 986         args = kzalloc(sizeof(*args), GFP_KERNEL);
 987         if (!args) {
 988                 sockfd_put(sock);
 989                 return -ENOMEM;
 990         }
 991
 992         for (i = 0; i < config->num_connections; i++) {
 993                 struct nbd_sock *nsock = config->socks[i];
 994
 995                 if (!nsock->dead)
 996                         continue;
 997
 998                 mutex_lock(&nsock->tx_lock);
 999                 if (!nsock->dead) {
1000                         mutex_unlock(&nsock->tx_lock);
1001                         continue;
1002                 }
1003                 sk_set_memalloc(sock->sk);
1004                 if (nbd->tag_set.timeout)
1005                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1006                 atomic_inc(&config->recv_threads);
1007                 refcount_inc(&nbd->config_refs);
1008                 old = nsock->sock;
1009                 nsock->fallback_index = -1;
1010                 nsock->sock = sock;
1011                 nsock->dead = false;
1012                 INIT_WORK(&args->work, recv_work);
1013                 args->index = i;
1014                 args->nbd = nbd;
1015                 nsock->cookie++;
1016                 mutex_unlock(&nsock->tx_lock);
1017                 sockfd_put(old);
1018
1019                 clear_bit(NBD_DISCONNECTED, &config->runtime_flags);
1020
1021                 /* We take the tx_mutex in an error path in the recv_work, so we
1022                  * need to queue_work outside of the tx_mutex.
1023                  */
1024                 queue_work(recv_workqueue, &args->work);
1025
1026                 atomic_inc(&config->live_connections);
1027                 wake_up(&config->conn_wait);
1028                 return 0;
1029         }
1030         sockfd_put(sock);
1031         kfree(args);
1032         return -ENOSPC;
1033 }
1034
1035 static void nbd_bdev_reset(struct block_device *bdev)
1036 {
1037         if (bdev->bd_openers > 1)
1038                 return;
1039         bd_set_size(bdev, 0);
1040 }
1041
1042 static void nbd_parse_flags(struct nbd_device *nbd)
1043 {
1044         struct nbd_config *config = nbd->config;
1045         if (config->flags & NBD_FLAG_READ_ONLY)
1046                 set_disk_ro(nbd->disk, true);
1047         else
1048                 set_disk_ro(nbd->disk, false);
1049         if (config->flags & NBD_FLAG_SEND_TRIM)
1050                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1051         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1052                 if (config->flags & NBD_FLAG_SEND_FUA)
1053                         blk_queue_write_cache(nbd->disk->queue, true, true);
1054                 else
1055                         blk_queue_write_cache(nbd->disk->queue, true, false);
1056         }
1057         else
1058                 blk_queue_write_cache(nbd->disk->queue, false, false);
1059 }
1060
1061 static void send_disconnects(struct nbd_device *nbd)
1062 {
1063         struct nbd_config *config = nbd->config;
1064         struct nbd_request request = {
1065                 .magic = htonl(NBD_REQUEST_MAGIC),
1066                 .type = htonl(NBD_CMD_DISC),
1067         };
1068         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1069         struct iov_iter from;
1070         int i, ret;
1071
1072         for (i = 0; i < config->num_connections; i++) {
1073                 struct nbd_sock *nsock = config->socks[i];
1074
1075                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1076                 mutex_lock(&nsock->tx_lock);
1077                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1078                 if (ret <= 0)
1079                         dev_err(disk_to_dev(nbd->disk),
1080                                 "Send disconnect failed %d\n", ret);
1081                 mutex_unlock(&nsock->tx_lock);
1082         }
1083 }
1084
1085 static int nbd_disconnect(struct nbd_device *nbd)
1086 {
1087         struct nbd_config *config = nbd->config;
1088
1089         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1090         set_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags);
1091         send_disconnects(nbd);
1092         return 0;
1093 }
1094
1095 static void nbd_clear_sock(struct nbd_device *nbd)
1096 {
1097         sock_shutdown(nbd);
1098         nbd_clear_que(nbd);
1099         nbd->task_setup = NULL;
1100 }
1101
1102 static void nbd_config_put(struct nbd_device *nbd)
1103 {
1104         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1105                                         &nbd->config_lock)) {
1106                 struct nbd_config *config = nbd->config;
1107                 nbd_dev_dbg_close(nbd);
1108                 nbd_size_clear(nbd);
1109                 if (test_and_clear_bit(NBD_HAS_PID_FILE,
1110                                        &config->runtime_flags))
1111                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1112                 nbd->task_recv = NULL;
1113                 nbd_clear_sock(nbd);
1114                 if (config->num_connections) {
1115                         int i;
1116                         for (i = 0; i < config->num_connections; i++) {
1117                                 sockfd_put(config->socks[i]->sock);
1118                                 kfree(config->socks[i]);
1119                         }
1120                         kfree(config->socks);
1121                 }
1122                 kfree(nbd->config);
1123                 nbd->config = NULL;
1124
1125                 nbd->tag_set.timeout = 0;
1126                 nbd->disk->queue->limits.discard_granularity = 0;
1127                 nbd->disk->queue->limits.discard_alignment = 0;
1128                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1129                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1130
1131                 mutex_unlock(&nbd->config_lock);
1132                 nbd_put(nbd);
1133                 module_put(THIS_MODULE);
1134         }
1135 }
1136
1137 static int nbd_start_device(struct nbd_device *nbd)
1138 {
1139         struct nbd_config *config = nbd->config;
1140         int num_connections = config->num_connections;
1141         int error = 0, i;
1142
1143         if (nbd->task_recv)
1144                 return -EBUSY;
1145         if (!config->socks)
1146                 return -EINVAL;
1147         if (num_connections > 1 &&
1148             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1149                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1150                 return -EINVAL;
1151         }
1152
1153         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1154         nbd->task_recv = current;
1155
1156         nbd_parse_flags(nbd);
1157
1158         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1159         if (error) {
1160                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1161                 return error;
1162         }
1163         set_bit(NBD_HAS_PID_FILE, &config->runtime_flags);
1164
1165         nbd_dev_dbg_init(nbd);
1166         for (i = 0; i < num_connections; i++) {
1167                 struct recv_thread_args *args;
1168
1169                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1170                 if (!args) {
1171                         sock_shutdown(nbd);
1172                         return -ENOMEM;
1173                 }
1174                 sk_set_memalloc(config->socks[i]->sock->sk);
1175                 if (nbd->tag_set.timeout)
1176                         config->socks[i]->sock->sk->sk_sndtimeo =
1177                                 nbd->tag_set.timeout;
1178                 atomic_inc(&config->recv_threads);
1179                 refcount_inc(&nbd->config_refs);
1180                 INIT_WORK(&args->work, recv_work);
1181                 args->nbd = nbd;
1182                 args->index = i;
1183                 queue_work(recv_workqueue, &args->work);
1184         }
1185         nbd_size_update(nbd);
1186         return error;
1187 }
1188
1189 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1190 {
1191         struct nbd_config *config = nbd->config;
1192         int ret;
1193
1194         ret = nbd_start_device(nbd);
1195         if (ret)
1196                 return ret;
1197
1198         if (max_part)
1199                 bdev->bd_invalidated = 1;
1200         mutex_unlock(&nbd->config_lock);
1201         ret = wait_event_interruptible(config->recv_wq,
1202                                          atomic_read(&config->recv_threads) == 0);
1203         if (ret)
1204                 sock_shutdown(nbd);
1205         mutex_lock(&nbd->config_lock);
1206         nbd_bdev_reset(bdev);
1207         /* user requested, ignore socket errors */
1208         if (test_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags))
1209                 ret = 0;
1210         if (test_bit(NBD_TIMEDOUT, &config->runtime_flags))
1211                 ret = -ETIMEDOUT;
1212         return ret;
1213 }
1214
1215 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1216                                  struct block_device *bdev)
1217 {
1218         sock_shutdown(nbd);
1219         kill_bdev(bdev);
1220         nbd_bdev_reset(bdev);
1221         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1222                                &nbd->config->runtime_flags))
1223                 nbd_config_put(nbd);
1224 }
1225
1226 /* Must be called with config_lock held */
1227 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1228                        unsigned int cmd, unsigned long arg)
1229 {
1230         struct nbd_config *config = nbd->config;
1231
1232         switch (cmd) {
1233         case NBD_DISCONNECT:
1234                 return nbd_disconnect(nbd);
1235         case NBD_CLEAR_SOCK:
1236                 nbd_clear_sock_ioctl(nbd, bdev);
1237                 return 0;
1238         case NBD_SET_SOCK:
1239                 return nbd_add_socket(nbd, arg, false);
1240         case NBD_SET_BLKSIZE:
1241                 if (!arg || !is_power_of_2(arg) || arg < 512 ||
1242                     arg > PAGE_SIZE)
1243                         return -EINVAL;
1244                 nbd_size_set(nbd, arg,
1245                              div_s64(config->bytesize, arg));
1246                 return 0;
1247         case NBD_SET_SIZE:
1248                 nbd_size_set(nbd, config->blksize,
1249                              div_s64(arg, config->blksize));
1250                 return 0;
1251         case NBD_SET_SIZE_BLOCKS:
1252                 nbd_size_set(nbd, config->blksize, arg);
1253                 return 0;
1254         case NBD_SET_TIMEOUT:
1255                 if (arg) {
1256                         nbd->tag_set.timeout = arg * HZ;
1257                         blk_queue_rq_timeout(nbd->disk->queue, arg * HZ);
1258                 }
1259                 return 0;
1260
1261         case NBD_SET_FLAGS:
1262                 config->flags = arg;
1263                 return 0;
1264         case NBD_DO_IT:
1265                 return nbd_start_device_ioctl(nbd, bdev);
1266         case NBD_CLEAR_QUE:
1267                 /*
1268                  * This is for compatibility only.  The queue is always cleared
1269                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1270                  */
1271                 return 0;
1272         case NBD_PRINT_DEBUG:
1273                 /*
1274                  * For compatibility only, we no longer keep a list of
1275                  * outstanding requests.
1276                  */
1277                 return 0;
1278         }
1279         return -ENOTTY;
1280 }
1281
1282 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1283                      unsigned int cmd, unsigned long arg)
1284 {
1285         struct nbd_device *nbd = bdev->bd_disk->private_data;
1286         struct nbd_config *config = nbd->config;
1287         int error = -EINVAL;
1288
1289         if (!capable(CAP_SYS_ADMIN))
1290                 return -EPERM;
1291
1292         /* The block layer will pass back some non-nbd ioctls in case we have
1293          * special handling for them, but we don't so just return an error.
1294          */
1295         if (_IOC_TYPE(cmd) != 0xab)
1296                 return -EINVAL;
1297
1298         mutex_lock(&nbd->config_lock);
1299
1300         /* Don't allow ioctl operations on a nbd device that was created with
1301          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1302          */
1303         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1304             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1305                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1306         else
1307                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1308         mutex_unlock(&nbd->config_lock);
1309         return error;
1310 }
1311
1312 static struct nbd_config *nbd_alloc_config(void)
1313 {
1314         struct nbd_config *config;
1315
1316         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1317         if (!config)
1318                 return NULL;
1319         atomic_set(&config->recv_threads, 0);
1320         init_waitqueue_head(&config->recv_wq);
1321         init_waitqueue_head(&config->conn_wait);
1322         config->blksize = 1024;
1323         atomic_set(&config->live_connections, 0);
1324         try_module_get(THIS_MODULE);
1325         return config;
1326 }
1327
1328 static int nbd_open(struct block_device *bdev, fmode_t mode)
1329 {
1330         struct nbd_device *nbd;
1331         int ret = 0;
1332
1333         mutex_lock(&nbd_index_mutex);
1334         nbd = bdev->bd_disk->private_data;
1335         if (!nbd) {
1336                 ret = -ENXIO;
1337                 goto out;
1338         }
1339         if (!refcount_inc_not_zero(&nbd->refs)) {
1340                 ret = -ENXIO;
1341                 goto out;
1342         }
1343         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1344                 struct nbd_config *config;
1345
1346                 mutex_lock(&nbd->config_lock);
1347                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1348                         mutex_unlock(&nbd->config_lock);
1349                         goto out;
1350                 }
1351                 config = nbd->config = nbd_alloc_config();
1352                 if (!config) {
1353                         ret = -ENOMEM;
1354                         mutex_unlock(&nbd->config_lock);
1355                         goto out;
1356                 }
1357                 refcount_set(&nbd->config_refs, 1);
1358                 refcount_inc(&nbd->refs);
1359                 mutex_unlock(&nbd->config_lock);
1360                 bdev->bd_invalidated = 1;
1361         } else if (nbd_disconnected(nbd->config)) {
1362                 bdev->bd_invalidated = 1;
1363         }
1364 out:
1365         mutex_unlock(&nbd_index_mutex);
1366         return ret;
1367 }
1368
1369 static void nbd_release(struct gendisk *disk, fmode_t mode)
1370 {
1371         struct nbd_device *nbd = disk->private_data;
1372         struct block_device *bdev = bdget_disk(disk, 0);
1373
1374         if (test_bit(NBD_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1375                         bdev->bd_openers == 0)
1376                 nbd_disconnect_and_put(nbd);
1377
1378         nbd_config_put(nbd);
1379         nbd_put(nbd);
1380 }
1381
1382 static const struct block_device_operations nbd_fops =
1383 {
1384         .owner =        THIS_MODULE,
1385         .open =         nbd_open,
1386         .release =      nbd_release,
1387         .ioctl =        nbd_ioctl,
1388         .compat_ioctl = nbd_ioctl,
1389 };
1390
1391 #if IS_ENABLED(CONFIG_DEBUG_FS)
1392
1393 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1394 {
1395         struct nbd_device *nbd = s->private;
1396
1397         if (nbd->task_recv)
1398                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1399
1400         return 0;
1401 }
1402
1403 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1404 {
1405         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1406 }
1407
1408 static const struct file_operations nbd_dbg_tasks_ops = {
1409         .open = nbd_dbg_tasks_open,
1410         .read = seq_read,
1411         .llseek = seq_lseek,
1412         .release = single_release,
1413 };
1414
1415 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1416 {
1417         struct nbd_device *nbd = s->private;
1418         u32 flags = nbd->config->flags;
1419
1420         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1421
1422         seq_puts(s, "Known flags:\n");
1423
1424         if (flags & NBD_FLAG_HAS_FLAGS)
1425                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1426         if (flags & NBD_FLAG_READ_ONLY)
1427                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1428         if (flags & NBD_FLAG_SEND_FLUSH)
1429                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1430         if (flags & NBD_FLAG_SEND_FUA)
1431                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1432         if (flags & NBD_FLAG_SEND_TRIM)
1433                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1434
1435         return 0;
1436 }
1437
1438 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1439 {
1440         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1441 }
1442
1443 static const struct file_operations nbd_dbg_flags_ops = {
1444         .open = nbd_dbg_flags_open,
1445         .read = seq_read,
1446         .llseek = seq_lseek,
1447         .release = single_release,
1448 };
1449
1450 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1451 {
1452         struct dentry *dir;
1453         struct nbd_config *config = nbd->config;
1454
1455         if (!nbd_dbg_dir)
1456                 return -EIO;
1457
1458         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1459         if (!dir) {
1460                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1461                         nbd_name(nbd));
1462                 return -EIO;
1463         }
1464         config->dbg_dir = dir;
1465
1466         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1467         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1468         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1469         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1470         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1471
1472         return 0;
1473 }
1474
1475 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1476 {
1477         debugfs_remove_recursive(nbd->config->dbg_dir);
1478 }
1479
1480 static int nbd_dbg_init(void)
1481 {
1482         struct dentry *dbg_dir;
1483
1484         dbg_dir = debugfs_create_dir("nbd", NULL);
1485         if (!dbg_dir)
1486                 return -EIO;
1487
1488         nbd_dbg_dir = dbg_dir;
1489
1490         return 0;
1491 }
1492
1493 static void nbd_dbg_close(void)
1494 {
1495         debugfs_remove_recursive(nbd_dbg_dir);
1496 }
1497
1498 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1499
1500 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1501 {
1502         return 0;
1503 }
1504
1505 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1506 {
1507 }
1508
1509 static int nbd_dbg_init(void)
1510 {
1511         return 0;
1512 }
1513
1514 static void nbd_dbg_close(void)
1515 {
1516 }
1517
1518 #endif
1519
1520 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1521                             unsigned int hctx_idx, unsigned int numa_node)
1522 {
1523         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1524         cmd->nbd = set->driver_data;
1525         cmd->flags = 0;
1526         mutex_init(&cmd->lock);
1527         return 0;
1528 }
1529
1530 static const struct blk_mq_ops nbd_mq_ops = {
1531         .queue_rq       = nbd_queue_rq,
1532         .complete       = nbd_complete_rq,
1533         .init_request   = nbd_init_request,
1534         .timeout        = nbd_xmit_timeout,
1535 };
1536
1537 static int nbd_dev_add(int index)
1538 {
1539         struct nbd_device *nbd;
1540         struct gendisk *disk;
1541         struct request_queue *q;
1542         int err = -ENOMEM;
1543
1544         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1545         if (!nbd)
1546                 goto out;
1547
1548         disk = alloc_disk(1 << part_shift);
1549         if (!disk)
1550                 goto out_free_nbd;
1551
1552         if (index >= 0) {
1553                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1554                                 GFP_KERNEL);
1555                 if (err == -ENOSPC)
1556                         err = -EEXIST;
1557         } else {
1558                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1559                 if (err >= 0)
1560                         index = err;
1561         }
1562         if (err < 0)
1563                 goto out_free_disk;
1564
1565         nbd->index = index;
1566         nbd->disk = disk;
1567         nbd->tag_set.ops = &nbd_mq_ops;
1568         nbd->tag_set.nr_hw_queues = 1;
1569         nbd->tag_set.queue_depth = 128;
1570         nbd->tag_set.numa_node = NUMA_NO_NODE;
1571         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1572         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1573                 BLK_MQ_F_SG_MERGE | BLK_MQ_F_BLOCKING;
1574         nbd->tag_set.driver_data = nbd;
1575
1576         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1577         if (err)
1578                 goto out_free_idr;
1579
1580         q = blk_mq_init_queue(&nbd->tag_set);
1581         if (IS_ERR(q)) {
1582                 err = PTR_ERR(q);
1583                 goto out_free_tags;
1584         }
1585         disk->queue = q;
1586
1587         /*
1588          * Tell the block layer that we are not a rotational device
1589          */
1590         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1591         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1592         disk->queue->limits.discard_granularity = 0;
1593         disk->queue->limits.discard_alignment = 0;
1594         blk_queue_max_discard_sectors(disk->queue, 0);
1595         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1596         blk_queue_max_segments(disk->queue, USHRT_MAX);
1597         blk_queue_max_hw_sectors(disk->queue, 65536);
1598         disk->queue->limits.max_sectors = 256;
1599
1600         mutex_init(&nbd->config_lock);
1601         refcount_set(&nbd->config_refs, 0);
1602         refcount_set(&nbd->refs, 1);
1603         INIT_LIST_HEAD(&nbd->list);
1604         disk->major = NBD_MAJOR;
1605         disk->first_minor = index << part_shift;
1606         disk->fops = &nbd_fops;
1607         disk->private_data = nbd;
1608         sprintf(disk->disk_name, "nbd%d", index);
1609         add_disk(disk);
1610         nbd_total_devices++;
1611         return index;
1612
1613 out_free_tags:
1614         blk_mq_free_tag_set(&nbd->tag_set);
1615 out_free_idr:
1616         idr_remove(&nbd_index_idr, index);
1617 out_free_disk:
1618         put_disk(disk);
1619 out_free_nbd:
1620         kfree(nbd);
1621 out:
1622         return err;
1623 }
1624
1625 static int find_free_cb(int id, void *ptr, void *data)
1626 {
1627         struct nbd_device *nbd = ptr;
1628         struct nbd_device **found = data;
1629
1630         if (!refcount_read(&nbd->config_refs)) {
1631                 *found = nbd;
1632                 return 1;
1633         }
1634         return 0;
1635 }
1636
1637 /* Netlink interface. */
1638 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1639         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1640         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1641         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1642         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1643         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1644         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1645         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1646         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1647         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1648 };
1649
1650 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1651         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1652 };
1653
1654 /* We don't use this right now since we don't parse the incoming list, but we
1655  * still want it here so userspace knows what to expect.
1656  */
1657 static const struct nla_policy __attribute__((unused))
1658 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1659         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1660         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1661 };
1662
1663 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1664 {
1665         struct nbd_device *nbd = NULL;
1666         struct nbd_config *config;
1667         int index = -1;
1668         int ret;
1669         bool put_dev = false;
1670
1671         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1672                 return -EPERM;
1673
1674         if (info->attrs[NBD_ATTR_INDEX])
1675                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1676         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1677                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1678                 return -EINVAL;
1679         }
1680         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1681                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1682                 return -EINVAL;
1683         }
1684 again:
1685         mutex_lock(&nbd_index_mutex);
1686         if (index == -1) {
1687                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1688                 if (ret == 0) {
1689                         int new_index;
1690                         new_index = nbd_dev_add(-1);
1691                         if (new_index < 0) {
1692                                 mutex_unlock(&nbd_index_mutex);
1693                                 printk(KERN_ERR "nbd: failed to add new device\n");
1694                                 return new_index;
1695                         }
1696                         nbd = idr_find(&nbd_index_idr, new_index);
1697                 }
1698         } else {
1699                 nbd = idr_find(&nbd_index_idr, index);
1700                 if (!nbd) {
1701                         ret = nbd_dev_add(index);
1702                         if (ret < 0) {
1703                                 mutex_unlock(&nbd_index_mutex);
1704                                 printk(KERN_ERR "nbd: failed to add new device\n");
1705                                 return ret;
1706                         }
1707                         nbd = idr_find(&nbd_index_idr, index);
1708                 }
1709         }
1710         if (!nbd) {
1711                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1712                        index);
1713                 mutex_unlock(&nbd_index_mutex);
1714                 return -EINVAL;
1715         }
1716         if (!refcount_inc_not_zero(&nbd->refs)) {
1717                 mutex_unlock(&nbd_index_mutex);
1718                 if (index == -1)
1719                         goto again;
1720                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1721                        index);
1722                 return -EINVAL;
1723         }
1724         mutex_unlock(&nbd_index_mutex);
1725
1726         mutex_lock(&nbd->config_lock);
1727         if (refcount_read(&nbd->config_refs)) {
1728                 mutex_unlock(&nbd->config_lock);
1729                 nbd_put(nbd);
1730                 if (index == -1)
1731                         goto again;
1732                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1733                 return -EBUSY;
1734         }
1735         if (WARN_ON(nbd->config)) {
1736                 mutex_unlock(&nbd->config_lock);
1737                 nbd_put(nbd);
1738                 return -EINVAL;
1739         }
1740         config = nbd->config = nbd_alloc_config();
1741         if (!nbd->config) {
1742                 mutex_unlock(&nbd->config_lock);
1743                 nbd_put(nbd);
1744                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1745                 return -ENOMEM;
1746         }
1747         refcount_set(&nbd->config_refs, 1);
1748         set_bit(NBD_BOUND, &config->runtime_flags);
1749
1750         if (info->attrs[NBD_ATTR_SIZE_BYTES]) {
1751                 u64 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1752                 nbd_size_set(nbd, config->blksize,
1753                              div64_u64(bytes, config->blksize));
1754         }
1755         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1756                 u64 bsize =
1757                         nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1758                 nbd_size_set(nbd, bsize, div64_u64(config->bytesize, bsize));
1759         }
1760         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1761                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1762                 nbd->tag_set.timeout = timeout * HZ;
1763                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1764         }
1765         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1766                 config->dead_conn_timeout =
1767                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1768                 config->dead_conn_timeout *= HZ;
1769         }
1770         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1771                 config->flags =
1772                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1773         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1774                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1775                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1776                         set_bit(NBD_DESTROY_ON_DISCONNECT,
1777                                 &config->runtime_flags);
1778                         put_dev = true;
1779                 }
1780                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1781                         set_bit(NBD_DISCONNECT_ON_CLOSE,
1782                                 &config->runtime_flags);
1783                 }
1784         }
1785
1786         if (info->attrs[NBD_ATTR_SOCKETS]) {
1787                 struct nlattr *attr;
1788                 int rem, fd;
1789
1790                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1791                                     rem) {
1792                         struct nlattr *socks[NBD_SOCK_MAX+1];
1793
1794                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1795                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1796                                 ret = -EINVAL;
1797                                 goto out;
1798                         }
1799                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1800                                                nbd_sock_policy, info->extack);
1801                         if (ret != 0) {
1802                                 printk(KERN_ERR "nbd: error processing sock list\n");
1803                                 ret = -EINVAL;
1804                                 goto out;
1805                         }
1806                         if (!socks[NBD_SOCK_FD])
1807                                 continue;
1808                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1809                         ret = nbd_add_socket(nbd, fd, true);
1810                         if (ret)
1811                                 goto out;
1812                 }
1813         }
1814         ret = nbd_start_device(nbd);
1815 out:
1816         mutex_unlock(&nbd->config_lock);
1817         if (!ret) {
1818                 set_bit(NBD_HAS_CONFIG_REF, &config->runtime_flags);
1819                 refcount_inc(&nbd->config_refs);
1820                 nbd_connect_reply(info, nbd->index);
1821         }
1822         nbd_config_put(nbd);
1823         if (put_dev)
1824                 nbd_put(nbd);
1825         return ret;
1826 }
1827
1828 static void nbd_disconnect_and_put(struct nbd_device *nbd)
1829 {
1830         mutex_lock(&nbd->config_lock);
1831         nbd_disconnect(nbd);
1832         nbd_clear_sock(nbd);
1833         mutex_unlock(&nbd->config_lock);
1834         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1835                                &nbd->config->runtime_flags))
1836                 nbd_config_put(nbd);
1837 }
1838
1839 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
1840 {
1841         struct nbd_device *nbd;
1842         int index;
1843
1844         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1845                 return -EPERM;
1846
1847         if (!info->attrs[NBD_ATTR_INDEX]) {
1848                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
1849                 return -EINVAL;
1850         }
1851         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1852         mutex_lock(&nbd_index_mutex);
1853         nbd = idr_find(&nbd_index_idr, index);
1854         if (!nbd) {
1855                 mutex_unlock(&nbd_index_mutex);
1856                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1857                        index);
1858                 return -EINVAL;
1859         }
1860         if (!refcount_inc_not_zero(&nbd->refs)) {
1861                 mutex_unlock(&nbd_index_mutex);
1862                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1863                        index);
1864                 return -EINVAL;
1865         }
1866         mutex_unlock(&nbd_index_mutex);
1867         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1868                 nbd_put(nbd);
1869                 return 0;
1870         }
1871         nbd_disconnect_and_put(nbd);
1872         nbd_config_put(nbd);
1873         nbd_put(nbd);
1874         return 0;
1875 }
1876
1877 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
1878 {
1879         struct nbd_device *nbd = NULL;
1880         struct nbd_config *config;
1881         int index;
1882         int ret = 0;
1883         bool put_dev = false;
1884
1885         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1886                 return -EPERM;
1887
1888         if (!info->attrs[NBD_ATTR_INDEX]) {
1889                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
1890                 return -EINVAL;
1891         }
1892         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1893         mutex_lock(&nbd_index_mutex);
1894         nbd = idr_find(&nbd_index_idr, index);
1895         if (!nbd) {
1896                 mutex_unlock(&nbd_index_mutex);
1897                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
1898                        index);
1899                 return -EINVAL;
1900         }
1901         if (!refcount_inc_not_zero(&nbd->refs)) {
1902                 mutex_unlock(&nbd_index_mutex);
1903                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1904                        index);
1905                 return -EINVAL;
1906         }
1907         mutex_unlock(&nbd_index_mutex);
1908
1909         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1910                 dev_err(nbd_to_dev(nbd),
1911                         "not configured, cannot reconfigure\n");
1912                 nbd_put(nbd);
1913                 return -EINVAL;
1914         }
1915
1916         mutex_lock(&nbd->config_lock);
1917         config = nbd->config;
1918         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1919             !nbd->task_recv) {
1920                 dev_err(nbd_to_dev(nbd),
1921                         "not configured, cannot reconfigure\n");
1922                 ret = -EINVAL;
1923                 goto out;
1924         }
1925
1926         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1927                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1928                 nbd->tag_set.timeout = timeout * HZ;
1929                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1930         }
1931         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1932                 config->dead_conn_timeout =
1933                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1934                 config->dead_conn_timeout *= HZ;
1935         }
1936         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1937                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1938                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1939                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1940                                               &config->runtime_flags))
1941                                 put_dev = true;
1942                 } else {
1943                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1944                                                &config->runtime_flags))
1945                                 refcount_inc(&nbd->refs);
1946                 }
1947
1948                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1949                         set_bit(NBD_DISCONNECT_ON_CLOSE,
1950                                         &config->runtime_flags);
1951                 } else {
1952                         clear_bit(NBD_DISCONNECT_ON_CLOSE,
1953                                         &config->runtime_flags);
1954                 }
1955         }
1956
1957         if (info->attrs[NBD_ATTR_SOCKETS]) {
1958                 struct nlattr *attr;
1959                 int rem, fd;
1960
1961                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1962                                     rem) {
1963                         struct nlattr *socks[NBD_SOCK_MAX+1];
1964
1965                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1966                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1967                                 ret = -EINVAL;
1968                                 goto out;
1969                         }
1970                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1971                                                nbd_sock_policy, info->extack);
1972                         if (ret != 0) {
1973                                 printk(KERN_ERR "nbd: error processing sock list\n");
1974                                 ret = -EINVAL;
1975                                 goto out;
1976                         }
1977                         if (!socks[NBD_SOCK_FD])
1978                                 continue;
1979                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1980                         ret = nbd_reconnect_socket(nbd, fd);
1981                         if (ret) {
1982                                 if (ret == -ENOSPC)
1983                                         ret = 0;
1984                                 goto out;
1985                         }
1986                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
1987                 }
1988         }
1989 out:
1990         mutex_unlock(&nbd->config_lock);
1991         nbd_config_put(nbd);
1992         nbd_put(nbd);
1993         if (put_dev)
1994                 nbd_put(nbd);
1995         return ret;
1996 }
1997
1998 static const struct genl_ops nbd_connect_genl_ops[] = {
1999         {
2000                 .cmd    = NBD_CMD_CONNECT,
2001                 .policy = nbd_attr_policy,
2002                 .doit   = nbd_genl_connect,
2003         },
2004         {
2005                 .cmd    = NBD_CMD_DISCONNECT,
2006                 .policy = nbd_attr_policy,
2007                 .doit   = nbd_genl_disconnect,
2008         },
2009         {
2010                 .cmd    = NBD_CMD_RECONFIGURE,
2011                 .policy = nbd_attr_policy,
2012                 .doit   = nbd_genl_reconfigure,
2013         },
2014         {
2015                 .cmd    = NBD_CMD_STATUS,
2016                 .policy = nbd_attr_policy,
2017                 .doit   = nbd_genl_status,
2018         },
2019 };
2020
2021 static const struct genl_multicast_group nbd_mcast_grps[] = {
2022         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2023 };
2024
2025 static struct genl_family nbd_genl_family __ro_after_init = {
2026         .hdrsize        = 0,
2027         .name           = NBD_GENL_FAMILY_NAME,
2028         .version        = NBD_GENL_VERSION,
2029         .module         = THIS_MODULE,
2030         .ops            = nbd_connect_genl_ops,
2031         .n_ops          = ARRAY_SIZE(nbd_connect_genl_ops),
2032         .maxattr        = NBD_ATTR_MAX,
2033         .mcgrps         = nbd_mcast_grps,
2034         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2035 };
2036
2037 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2038 {
2039         struct nlattr *dev_opt;
2040         u8 connected = 0;
2041         int ret;
2042
2043         /* This is a little racey, but for status it's ok.  The
2044          * reason we don't take a ref here is because we can't
2045          * take a ref in the index == -1 case as we would need
2046          * to put under the nbd_index_mutex, which could
2047          * deadlock if we are configured to remove ourselves
2048          * once we're disconnected.
2049          */
2050         if (refcount_read(&nbd->config_refs))
2051                 connected = 1;
2052         dev_opt = nla_nest_start(reply, NBD_DEVICE_ITEM);
2053         if (!dev_opt)
2054                 return -EMSGSIZE;
2055         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2056         if (ret)
2057                 return -EMSGSIZE;
2058         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2059                          connected);
2060         if (ret)
2061                 return -EMSGSIZE;
2062         nla_nest_end(reply, dev_opt);
2063         return 0;
2064 }
2065
2066 static int status_cb(int id, void *ptr, void *data)
2067 {
2068         struct nbd_device *nbd = ptr;
2069         return populate_nbd_status(nbd, (struct sk_buff *)data);
2070 }
2071
2072 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2073 {
2074         struct nlattr *dev_list;
2075         struct sk_buff *reply;
2076         void *reply_head;
2077         size_t msg_size;
2078         int index = -1;
2079         int ret = -ENOMEM;
2080
2081         if (info->attrs[NBD_ATTR_INDEX])
2082                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2083
2084         mutex_lock(&nbd_index_mutex);
2085
2086         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2087                                   nla_attr_size(sizeof(u8)));
2088         msg_size *= (index == -1) ? nbd_total_devices : 1;
2089
2090         reply = genlmsg_new(msg_size, GFP_KERNEL);
2091         if (!reply)
2092                 goto out;
2093         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2094                                        NBD_CMD_STATUS);
2095         if (!reply_head) {
2096                 nlmsg_free(reply);
2097                 goto out;
2098         }
2099
2100         dev_list = nla_nest_start(reply, NBD_ATTR_DEVICE_LIST);
2101         if (index == -1) {
2102                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2103                 if (ret) {
2104                         nlmsg_free(reply);
2105                         goto out;
2106                 }
2107         } else {
2108                 struct nbd_device *nbd;
2109                 nbd = idr_find(&nbd_index_idr, index);
2110                 if (nbd) {
2111                         ret = populate_nbd_status(nbd, reply);
2112                         if (ret) {
2113                                 nlmsg_free(reply);
2114                                 goto out;
2115                         }
2116                 }
2117         }
2118         nla_nest_end(reply, dev_list);
2119         genlmsg_end(reply, reply_head);
2120         genlmsg_reply(reply, info);
2121         ret = 0;
2122 out:
2123         mutex_unlock(&nbd_index_mutex);
2124         return ret;
2125 }
2126
2127 static void nbd_connect_reply(struct genl_info *info, int index)
2128 {
2129         struct sk_buff *skb;
2130         void *msg_head;
2131         int ret;
2132
2133         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2134         if (!skb)
2135                 return;
2136         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2137                                      NBD_CMD_CONNECT);
2138         if (!msg_head) {
2139                 nlmsg_free(skb);
2140                 return;
2141         }
2142         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2143         if (ret) {
2144                 nlmsg_free(skb);
2145                 return;
2146         }
2147         genlmsg_end(skb, msg_head);
2148         genlmsg_reply(skb, info);
2149 }
2150
2151 static void nbd_mcast_index(int index)
2152 {
2153         struct sk_buff *skb;
2154         void *msg_head;
2155         int ret;
2156
2157         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2158         if (!skb)
2159                 return;
2160         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2161                                      NBD_CMD_LINK_DEAD);
2162         if (!msg_head) {
2163                 nlmsg_free(skb);
2164                 return;
2165         }
2166         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2167         if (ret) {
2168                 nlmsg_free(skb);
2169                 return;
2170         }
2171         genlmsg_end(skb, msg_head);
2172         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2173 }
2174
2175 static void nbd_dead_link_work(struct work_struct *work)
2176 {
2177         struct link_dead_args *args = container_of(work, struct link_dead_args,
2178                                                    work);
2179         nbd_mcast_index(args->index);
2180         kfree(args);
2181 }
2182
2183 static int __init nbd_init(void)
2184 {
2185         int i;
2186
2187         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2188
2189         if (max_part < 0) {
2190                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2191                 return -EINVAL;
2192         }
2193
2194         part_shift = 0;
2195         if (max_part > 0) {
2196                 part_shift = fls(max_part);
2197
2198                 /*
2199                  * Adjust max_part according to part_shift as it is exported
2200                  * to user space so that user can know the max number of
2201                  * partition kernel should be able to manage.
2202                  *
2203                  * Note that -1 is required because partition 0 is reserved
2204                  * for the whole disk.
2205                  */
2206                 max_part = (1UL << part_shift) - 1;
2207         }
2208
2209         if ((1UL << part_shift) > DISK_MAX_PARTS)
2210                 return -EINVAL;
2211
2212         if (nbds_max > 1UL << (MINORBITS - part_shift))
2213                 return -EINVAL;
2214         recv_workqueue = alloc_workqueue("knbd-recv",
2215                                          WQ_MEM_RECLAIM | WQ_HIGHPRI |
2216                                          WQ_UNBOUND, 0);
2217         if (!recv_workqueue)
2218                 return -ENOMEM;
2219
2220         if (register_blkdev(NBD_MAJOR, "nbd")) {
2221                 destroy_workqueue(recv_workqueue);
2222                 return -EIO;
2223         }
2224
2225         if (genl_register_family(&nbd_genl_family)) {
2226                 unregister_blkdev(NBD_MAJOR, "nbd");
2227                 destroy_workqueue(recv_workqueue);
2228                 return -EINVAL;
2229         }
2230         nbd_dbg_init();
2231
2232         mutex_lock(&nbd_index_mutex);
2233         for (i = 0; i < nbds_max; i++)
2234                 nbd_dev_add(i);
2235         mutex_unlock(&nbd_index_mutex);
2236         return 0;
2237 }
2238
2239 static int nbd_exit_cb(int id, void *ptr, void *data)
2240 {
2241         struct list_head *list = (struct list_head *)data;
2242         struct nbd_device *nbd = ptr;
2243
2244         list_add_tail(&nbd->list, list);
2245         return 0;
2246 }
2247
2248 static void __exit nbd_cleanup(void)
2249 {
2250         struct nbd_device *nbd;
2251         LIST_HEAD(del_list);
2252
2253         nbd_dbg_close();
2254
2255         mutex_lock(&nbd_index_mutex);
2256         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2257         mutex_unlock(&nbd_index_mutex);
2258
2259         while (!list_empty(&del_list)) {
2260                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2261                 list_del_init(&nbd->list);
2262                 if (refcount_read(&nbd->refs) != 1)
2263                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2264                 nbd_put(nbd);
2265         }
2266
2267         idr_destroy(&nbd_index_idr);
2268         genl_unregister_family(&nbd_genl_family);
2269         destroy_workqueue(recv_workqueue);
2270         unregister_blkdev(NBD_MAJOR, "nbd");
2271 }
2272
2273 module_init(nbd_init);
2274 module_exit(nbd_cleanup);
2275
2276 MODULE_DESCRIPTION("Network Block Device");
2277 MODULE_LICENSE("GPL");
2278
2279 module_param(nbds_max, int, 0444);
2280 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2281 module_param(max_part, int, 0444);
2282 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");