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 #define pr_fmt(fmt) "nbd: " fmt
  15
  16 #include <linux/major.h>
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/module.h>
  20 #include <linux/init.h>
  21 #include <linux/sched.h>
  22 #include <linux/sched/mm.h>
  23 #include <linux/fs.h>
  24 #include <linux/bio.h>
  25 #include <linux/stat.h>
  26 #include <linux/errno.h>
  27 #include <linux/file.h>
  28 #include <linux/ioctl.h>
  29 #include <linux/mutex.h>
  30 #include <linux/compiler.h>
  31 #include <linux/completion.h>
  32 #include <linux/err.h>
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <net/sock.h>
  36 #include <linux/net.h>
  37 #include <linux/kthread.h>
  38 #include <linux/types.h>
  39 #include <linux/debugfs.h>
  40 #include <linux/blk-mq.h>
  41
  42 #include <linux/uaccess.h>
  43 #include <asm/types.h>
  44
  45 #include <linux/nbd.h>
  46 #include <linux/nbd-netlink.h>
  47 #include <net/genetlink.h>
  48
  49 #define CREATE_TRACE_POINTS
  50 #include <trace/events/nbd.h>
  51
  52 static DEFINE_IDR(nbd_index_idr);
  53 static DEFINE_MUTEX(nbd_index_mutex);
  54 static struct workqueue_struct *nbd_del_wq;
  55 static int nbd_total_devices = 0;
  56
  57 struct nbd_sock {
  58         struct socket *sock;
  59         struct mutex tx_lock;
  60         struct request *pending;
  61         int sent;
  62         bool dead;
  63         int fallback_index;
  64         int cookie;
  65 };
  66
  67 struct recv_thread_args {
  68         struct work_struct work;
  69         struct nbd_device *nbd;
  70         int index;
  71 };
  72
  73 struct link_dead_args {
  74         struct work_struct work;
  75         int index;
  76 };
  77
  78 #define NBD_RT_TIMEDOUT                 0
  79 #define NBD_RT_DISCONNECT_REQUESTED     1
  80 #define NBD_RT_DISCONNECTED             2
  81 #define NBD_RT_HAS_PID_FILE             3
  82 #define NBD_RT_HAS_CONFIG_REF           4
  83 #define NBD_RT_BOUND                    5
  84 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  85 #define NBD_RT_HAS_BACKEND_FILE         7
  86
  87 #define NBD_DESTROY_ON_DISCONNECT       0
  88 #define NBD_DISCONNECT_REQUESTED        1
  89
  90 struct nbd_config {
  91         u32 flags;
  92         unsigned long runtime_flags;
  93         u64 dead_conn_timeout;
  94
  95         struct nbd_sock **socks;
  96         int num_connections;
  97         atomic_t live_connections;
  98         wait_queue_head_t conn_wait;
  99
 100         atomic_t recv_threads;
 101         wait_queue_head_t recv_wq;
 102         unsigned int blksize_bits;
 103         loff_t bytesize;
 104 #if IS_ENABLED(CONFIG_DEBUG_FS)
 105         struct dentry *dbg_dir;
 106 #endif
 107 };
 108
 109 static inline unsigned int nbd_blksize(struct nbd_config *config)
 110 {
 111         return 1u << config->blksize_bits;
 112 }
 113
 114 struct nbd_device {
 115         struct blk_mq_tag_set tag_set;
 116
 117         int index;
 118         refcount_t config_refs;
 119         refcount_t refs;
 120         struct nbd_config *config;
 121         struct mutex config_lock;
 122         struct gendisk *disk;
 123         struct workqueue_struct *recv_workq;
 124         struct work_struct remove_work;
 125
 126         struct list_head list;
 127         struct task_struct *task_setup;
 128
 129         unsigned long flags;
 130         pid_t pid; /* pid of nbd-client, if attached */
 131
 132         char *backend;
 133 };
 134
 135 #define NBD_CMD_REQUEUED        1
 136 /*
 137  * This flag will be set if nbd_queue_rq() succeed, and will be checked and
 138  * cleared in completion. Both setting and clearing of the flag are protected
 139  * by cmd->lock.
 140  */
 141 #define NBD_CMD_INFLIGHT        2
 142
 143 struct nbd_cmd {
 144         struct nbd_device *nbd;
 145         struct mutex lock;
 146         int index;
 147         int cookie;
 148         int retries;
 149         blk_status_t status;
 150         unsigned long flags;
 151         u32 cmd_cookie;
 152 };
 153
 154 #if IS_ENABLED(CONFIG_DEBUG_FS)
 155 static struct dentry *nbd_dbg_dir;
 156 #endif
 157
 158 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 159
 160 #define NBD_MAGIC 0x68797548
 161
 162 #define NBD_DEF_BLKSIZE_BITS 10
 163
 164 static unsigned int nbds_max = 16;
 165 static int max_part = 16;
 166 static int part_shift;
 167
 168 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 169 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 170 static void nbd_config_put(struct nbd_device *nbd);
 171 static void nbd_connect_reply(struct genl_info *info, int index);
 172 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 173 static void nbd_dead_link_work(struct work_struct *work);
 174 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 175
 176 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 177 {
 178         return disk_to_dev(nbd->disk);
 179 }
 180
 181 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 182 {
 183         struct request *req = blk_mq_rq_from_pdu(cmd);
 184
 185         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 186                 blk_mq_requeue_request(req, true);
 187 }
 188
 189 #define NBD_COOKIE_BITS 32
 190
 191 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 192 {
 193         struct request *req = blk_mq_rq_from_pdu(cmd);
 194         u32 tag = blk_mq_unique_tag(req);
 195         u64 cookie = cmd->cmd_cookie;
 196
 197         return (cookie << NBD_COOKIE_BITS) | tag;
 198 }
 199
 200 static u32 nbd_handle_to_tag(u64 handle)
 201 {
 202         return (u32)handle;
 203 }
 204
 205 static u32 nbd_handle_to_cookie(u64 handle)
 206 {
 207         return (u32)(handle >> NBD_COOKIE_BITS);
 208 }
 209
 210 static const char *nbdcmd_to_ascii(int cmd)
 211 {
 212         switch (cmd) {
 213         case  NBD_CMD_READ: return "read";
 214         case NBD_CMD_WRITE: return "write";
 215         case  NBD_CMD_DISC: return "disconnect";
 216         case NBD_CMD_FLUSH: return "flush";
 217         case  NBD_CMD_TRIM: return "trim/discard";
 218         }
 219         return "invalid";
 220 }
 221
 222 static ssize_t pid_show(struct device *dev,
 223                         struct device_attribute *attr, char *buf)
 224 {
 225         struct gendisk *disk = dev_to_disk(dev);
 226         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 227
 228         return sprintf(buf, "%d\n", nbd->pid);
 229 }
 230
 231 static const struct device_attribute pid_attr = {
 232         .attr = { .name = "pid", .mode = 0444},
 233         .show = pid_show,
 234 };
 235
 236 static ssize_t backend_show(struct device *dev,
 237                 struct device_attribute *attr, char *buf)
 238 {
 239         struct gendisk *disk = dev_to_disk(dev);
 240         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 241
 242         return sprintf(buf, "%s\n", nbd->backend ?: "");
 243 }
 244
 245 static const struct device_attribute backend_attr = {
 246         .attr = { .name = "backend", .mode = 0444},
 247         .show = backend_show,
 248 };
 249
 250 static void nbd_dev_remove(struct nbd_device *nbd)
 251 {
 252         struct gendisk *disk = nbd->disk;
 253
 254         del_gendisk(disk);
 255         put_disk(disk);
 256         blk_mq_free_tag_set(&nbd->tag_set);
 257
 258         /*
 259          * Remove from idr after del_gendisk() completes, so if the same ID is
 260          * reused, the following add_disk() will succeed.
 261          */
 262         mutex_lock(&nbd_index_mutex);
 263         idr_remove(&nbd_index_idr, nbd->index);
 264         mutex_unlock(&nbd_index_mutex);
 265         destroy_workqueue(nbd->recv_workq);
 266         kfree(nbd);
 267 }
 268
 269 static void nbd_dev_remove_work(struct work_struct *work)
 270 {
 271         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 272 }
 273
 274 static void nbd_put(struct nbd_device *nbd)
 275 {
 276         if (!refcount_dec_and_test(&nbd->refs))
 277                 return;
 278
 279         /* Call del_gendisk() asynchrounously to prevent deadlock */
 280         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 281                 queue_work(nbd_del_wq, &nbd->remove_work);
 282         else
 283                 nbd_dev_remove(nbd);
 284 }
 285
 286 static int nbd_disconnected(struct nbd_config *config)
 287 {
 288         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 289                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 290 }
 291
 292 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 293                                 int notify)
 294 {
 295         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 296                 struct link_dead_args *args;
 297                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 298                 if (args) {
 299                         INIT_WORK(&args->work, nbd_dead_link_work);
 300                         args->index = nbd->index;
 301                         queue_work(system_wq, &args->work);
 302                 }
 303         }
 304         if (!nsock->dead) {
 305                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 306                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 307                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 308                                                &nbd->config->runtime_flags)) {
 309                                 set_bit(NBD_RT_DISCONNECTED,
 310                                         &nbd->config->runtime_flags);
 311                                 dev_info(nbd_to_dev(nbd),
 312                                         "Disconnected due to user request.\n");
 313                         }
 314                 }
 315         }
 316         nsock->dead = true;
 317         nsock->pending = NULL;
 318         nsock->sent = 0;
 319 }
 320
 321 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 322                 loff_t blksize)
 323 {
 324         if (!blksize)
 325                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 326
 327         if (blk_validate_block_size(blksize))
 328                 return -EINVAL;
 329
 330         nbd->config->bytesize = bytesize;
 331         nbd->config->blksize_bits = __ffs(blksize);
 332
 333         if (!nbd->pid)
 334                 return 0;
 335
 336         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 337                 nbd->disk->queue->limits.discard_granularity = blksize;
 338                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 339         }
 340         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 341         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 342
 343         if (max_part)
 344                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 345         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 346                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 347         return 0;
 348 }
 349
 350 static void nbd_complete_rq(struct request *req)
 351 {
 352         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 353
 354         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 355                 cmd->status ? "failed" : "done");
 356
 357         blk_mq_end_request(req, cmd->status);
 358 }
 359
 360 /*
 361  * Forcibly shutdown the socket causing all listeners to error
 362  */
 363 static void sock_shutdown(struct nbd_device *nbd)
 364 {
 365         struct nbd_config *config = nbd->config;
 366         int i;
 367
 368         if (config->num_connections == 0)
 369                 return;
 370         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 371                 return;
 372
 373         for (i = 0; i < config->num_connections; i++) {
 374                 struct nbd_sock *nsock = config->socks[i];
 375                 mutex_lock(&nsock->tx_lock);
 376                 nbd_mark_nsock_dead(nbd, nsock, 0);
 377                 mutex_unlock(&nsock->tx_lock);
 378         }
 379         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 380 }
 381
 382 static u32 req_to_nbd_cmd_type(struct request *req)
 383 {
 384         switch (req_op(req)) {
 385         case REQ_OP_DISCARD:
 386                 return NBD_CMD_TRIM;
 387         case REQ_OP_FLUSH:
 388                 return NBD_CMD_FLUSH;
 389         case REQ_OP_WRITE:
 390                 return NBD_CMD_WRITE;
 391         case REQ_OP_READ:
 392                 return NBD_CMD_READ;
 393         default:
 394                 return U32_MAX;
 395         }
 396 }
 397
 398 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req)
 399 {
 400         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 401         struct nbd_device *nbd = cmd->nbd;
 402         struct nbd_config *config;
 403
 404         if (!mutex_trylock(&cmd->lock))
 405                 return BLK_EH_RESET_TIMER;
 406
 407         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 408                 mutex_unlock(&cmd->lock);
 409                 return BLK_EH_DONE;
 410         }
 411
 412         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 413                 cmd->status = BLK_STS_TIMEOUT;
 414                 __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 415                 mutex_unlock(&cmd->lock);
 416                 goto done;
 417         }
 418         config = nbd->config;
 419
 420         if (config->num_connections > 1 ||
 421             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 422                 dev_err_ratelimited(nbd_to_dev(nbd),
 423                                     "Connection timed out, retrying (%d/%d alive)\n",
 424                                     atomic_read(&config->live_connections),
 425                                     config->num_connections);
 426                 /*
 427                  * Hooray we have more connections, requeue this IO, the submit
 428                  * path will put it on a real connection. Or if only one
 429                  * connection is configured, the submit path will wait util
 430                  * a new connection is reconfigured or util dead timeout.
 431                  */
 432                 if (config->socks) {
 433                         if (cmd->index < config->num_connections) {
 434                                 struct nbd_sock *nsock =
 435                                         config->socks[cmd->index];
 436                                 mutex_lock(&nsock->tx_lock);
 437                                 /* We can have multiple outstanding requests, so
 438                                  * we don't want to mark the nsock dead if we've
 439                                  * already reconnected with a new socket, so
 440                                  * only mark it dead if its the same socket we
 441                                  * were sent out on.
 442                                  */
 443                                 if (cmd->cookie == nsock->cookie)
 444                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 445                                 mutex_unlock(&nsock->tx_lock);
 446                         }
 447                         mutex_unlock(&cmd->lock);
 448                         nbd_requeue_cmd(cmd);
 449                         nbd_config_put(nbd);
 450                         return BLK_EH_DONE;
 451                 }
 452         }
 453
 454         if (!nbd->tag_set.timeout) {
 455                 /*
 456                  * Userspace sets timeout=0 to disable socket disconnection,
 457                  * so just warn and reset the timer.
 458                  */
 459                 struct nbd_sock *nsock = config->socks[cmd->index];
 460                 cmd->retries++;
 461                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 462                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 463                         (unsigned long long)blk_rq_pos(req) << 9,
 464                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 465
 466                 mutex_lock(&nsock->tx_lock);
 467                 if (cmd->cookie != nsock->cookie) {
 468                         nbd_requeue_cmd(cmd);
 469                         mutex_unlock(&nsock->tx_lock);
 470                         mutex_unlock(&cmd->lock);
 471                         nbd_config_put(nbd);
 472                         return BLK_EH_DONE;
 473                 }
 474                 mutex_unlock(&nsock->tx_lock);
 475                 mutex_unlock(&cmd->lock);
 476                 nbd_config_put(nbd);
 477                 return BLK_EH_RESET_TIMER;
 478         }
 479
 480         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 481         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 482         cmd->status = BLK_STS_IOERR;
 483         __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 484         mutex_unlock(&cmd->lock);
 485         sock_shutdown(nbd);
 486         nbd_config_put(nbd);
 487 done:
 488         blk_mq_complete_request(req);
 489         return BLK_EH_DONE;
 490 }
 491
 492 /*
 493  *  Send or receive packet. Return a positive value on success and
 494  *  negtive value on failue, and never return 0.
 495  */
 496 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 497                      struct iov_iter *iter, int msg_flags, int *sent)
 498 {
 499         struct nbd_config *config = nbd->config;
 500         struct socket *sock = config->socks[index]->sock;
 501         int result;
 502         struct msghdr msg;
 503         unsigned int noreclaim_flag;
 504
 505         if (unlikely(!sock)) {
 506                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 507                         "Attempted %s on closed socket in sock_xmit\n",
 508                         (send ? "send" : "recv"));
 509                 return -EINVAL;
 510         }
 511
 512         msg.msg_iter = *iter;
 513
 514         noreclaim_flag = memalloc_noreclaim_save();
 515         do {
 516                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 517                 msg.msg_name = NULL;
 518                 msg.msg_namelen = 0;
 519                 msg.msg_control = NULL;
 520                 msg.msg_controllen = 0;
 521                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 522
 523                 if (send)
 524                         result = sock_sendmsg(sock, &msg);
 525                 else
 526                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 527
 528                 if (result <= 0) {
 529                         if (result == 0)
 530                                 result = -EPIPE; /* short read */
 531                         break;
 532                 }
 533                 if (sent)
 534                         *sent += result;
 535         } while (msg_data_left(&msg));
 536
 537         memalloc_noreclaim_restore(noreclaim_flag);
 538
 539         return result;
 540 }
 541
 542 /*
 543  * Different settings for sk->sk_sndtimeo can result in different return values
 544  * if there is a signal pending when we enter sendmsg, because reasons?
 545  */
 546 static inline int was_interrupted(int result)
 547 {
 548         return result == -ERESTARTSYS || result == -EINTR;
 549 }
 550
 551 /* always call with the tx_lock held */
 552 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 553 {
 554         struct request *req = blk_mq_rq_from_pdu(cmd);
 555         struct nbd_config *config = nbd->config;
 556         struct nbd_sock *nsock = config->socks[index];
 557         int result;
 558         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 559         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 560         struct iov_iter from;
 561         unsigned long size = blk_rq_bytes(req);
 562         struct bio *bio;
 563         u64 handle;
 564         u32 type;
 565         u32 nbd_cmd_flags = 0;
 566         int sent = nsock->sent, skip = 0;
 567
 568         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 569
 570         type = req_to_nbd_cmd_type(req);
 571         if (type == U32_MAX)
 572                 return -EIO;
 573
 574         if (rq_data_dir(req) == WRITE &&
 575             (config->flags & NBD_FLAG_READ_ONLY)) {
 576                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 577                                     "Write on read-only\n");
 578                 return -EIO;
 579         }
 580
 581         if (req->cmd_flags & REQ_FUA)
 582                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 583
 584         /* We did a partial send previously, and we at least sent the whole
 585          * request struct, so just go and send the rest of the pages in the
 586          * request.
 587          */
 588         if (sent) {
 589                 if (sent >= sizeof(request)) {
 590                         skip = sent - sizeof(request);
 591
 592                         /* initialize handle for tracing purposes */
 593                         handle = nbd_cmd_handle(cmd);
 594
 595                         goto send_pages;
 596                 }
 597                 iov_iter_advance(&from, sent);
 598         } else {
 599                 cmd->cmd_cookie++;
 600         }
 601         cmd->index = index;
 602         cmd->cookie = nsock->cookie;
 603         cmd->retries = 0;
 604         request.type = htonl(type | nbd_cmd_flags);
 605         if (type != NBD_CMD_FLUSH) {
 606                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 607                 request.len = htonl(size);
 608         }
 609         handle = nbd_cmd_handle(cmd);
 610         memcpy(request.handle, &handle, sizeof(handle));
 611
 612         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 613
 614         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 615                 req, nbdcmd_to_ascii(type),
 616                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 617         result = sock_xmit(nbd, index, 1, &from,
 618                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 619         trace_nbd_header_sent(req, handle);
 620         if (result < 0) {
 621                 if (was_interrupted(result)) {
 622                         /* If we havne't sent anything we can just return BUSY,
 623                          * however if we have sent something we need to make
 624                          * sure we only allow this req to be sent until we are
 625                          * completely done.
 626                          */
 627                         if (sent) {
 628                                 nsock->pending = req;
 629                                 nsock->sent = sent;
 630                         }
 631                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 632                         return BLK_STS_RESOURCE;
 633                 }
 634                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 635                         "Send control failed (result %d)\n", result);
 636                 return -EAGAIN;
 637         }
 638 send_pages:
 639         if (type != NBD_CMD_WRITE)
 640                 goto out;
 641
 642         bio = req->bio;
 643         while (bio) {
 644                 struct bio *next = bio->bi_next;
 645                 struct bvec_iter iter;
 646                 struct bio_vec bvec;
 647
 648                 bio_for_each_segment(bvec, bio, iter) {
 649                         bool is_last = !next && bio_iter_last(bvec, iter);
 650                         int flags = is_last ? 0 : MSG_MORE;
 651
 652                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 653                                 req, bvec.bv_len);
 654                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 655                         if (skip) {
 656                                 if (skip >= iov_iter_count(&from)) {
 657                                         skip -= iov_iter_count(&from);
 658                                         continue;
 659                                 }
 660                                 iov_iter_advance(&from, skip);
 661                                 skip = 0;
 662                         }
 663                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 664                         if (result < 0) {
 665                                 if (was_interrupted(result)) {
 666                                         /* We've already sent the header, we
 667                                          * have no choice but to set pending and
 668                                          * return BUSY.
 669                                          */
 670                                         nsock->pending = req;
 671                                         nsock->sent = sent;
 672                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 673                                         return BLK_STS_RESOURCE;
 674                                 }
 675                                 dev_err(disk_to_dev(nbd->disk),
 676                                         "Send data failed (result %d)\n",
 677                                         result);
 678                                 return -EAGAIN;
 679                         }
 680                         /*
 681                          * The completion might already have come in,
 682                          * so break for the last one instead of letting
 683                          * the iterator do it. This prevents use-after-free
 684                          * of the bio.
 685                          */
 686                         if (is_last)
 687                                 break;
 688                 }
 689                 bio = next;
 690         }
 691 out:
 692         trace_nbd_payload_sent(req, handle);
 693         nsock->pending = NULL;
 694         nsock->sent = 0;
 695         return 0;
 696 }
 697
 698 static int nbd_read_reply(struct nbd_device *nbd, int index,
 699                           struct nbd_reply *reply)
 700 {
 701         struct kvec iov = {.iov_base = reply, .iov_len = sizeof(*reply)};
 702         struct iov_iter to;
 703         int result;
 704
 705         reply->magic = 0;
 706         iov_iter_kvec(&to, READ, &iov, 1, sizeof(*reply));
 707         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 708         if (result < 0) {
 709                 if (!nbd_disconnected(nbd->config))
 710                         dev_err(disk_to_dev(nbd->disk),
 711                                 "Receive control failed (result %d)\n", result);
 712                 return result;
 713         }
 714
 715         if (ntohl(reply->magic) != NBD_REPLY_MAGIC) {
 716                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 717                                 (unsigned long)ntohl(reply->magic));
 718                 return -EPROTO;
 719         }
 720
 721         return 0;
 722 }
 723
 724 /* NULL returned = something went wrong, inform userspace */
 725 static struct nbd_cmd *nbd_handle_reply(struct nbd_device *nbd, int index,
 726                                         struct nbd_reply *reply)
 727 {
 728         int result;
 729         struct nbd_cmd *cmd;
 730         struct request *req = NULL;
 731         u64 handle;
 732         u16 hwq;
 733         u32 tag;
 734         int ret = 0;
 735
 736         memcpy(&handle, reply->handle, sizeof(handle));
 737         tag = nbd_handle_to_tag(handle);
 738         hwq = blk_mq_unique_tag_to_hwq(tag);
 739         if (hwq < nbd->tag_set.nr_hw_queues)
 740                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 741                                        blk_mq_unique_tag_to_tag(tag));
 742         if (!req || !blk_mq_request_started(req)) {
 743                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 744                         tag, req);
 745                 return ERR_PTR(-ENOENT);
 746         }
 747         trace_nbd_header_received(req, handle);
 748         cmd = blk_mq_rq_to_pdu(req);
 749
 750         mutex_lock(&cmd->lock);
 751         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 752                 dev_err(disk_to_dev(nbd->disk), "Suspicious reply %d (status %u flags %lu)",
 753                         tag, cmd->status, cmd->flags);
 754                 ret = -ENOENT;
 755                 goto out;
 756         }
 757         if (cmd->index != index) {
 758                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d from different sock %d (expected %d)",
 759                         tag, index, cmd->index);
 760                 ret = -ENOENT;
 761                 goto out;
 762         }
 763         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 764                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 765                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 766                 ret = -ENOENT;
 767                 goto out;
 768         }
 769         if (cmd->status != BLK_STS_OK) {
 770                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 771                         req);
 772                 ret = -ENOENT;
 773                 goto out;
 774         }
 775         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 776                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 777                         req);
 778                 ret = -ENOENT;
 779                 goto out;
 780         }
 781         if (ntohl(reply->error)) {
 782                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 783                         ntohl(reply->error));
 784                 cmd->status = BLK_STS_IOERR;
 785                 goto out;
 786         }
 787
 788         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 789         if (rq_data_dir(req) != WRITE) {
 790                 struct req_iterator iter;
 791                 struct bio_vec bvec;
 792                 struct iov_iter to;
 793
 794                 rq_for_each_segment(bvec, req, iter) {
 795                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 796                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 797                         if (result < 0) {
 798                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 799                                         result);
 800                                 /*
 801                                  * If we've disconnected, we need to make sure we
 802                                  * complete this request, otherwise error out
 803                                  * and let the timeout stuff handle resubmitting
 804                                  * this request onto another connection.
 805                                  */
 806                                 if (nbd_disconnected(nbd->config)) {
 807                                         cmd->status = BLK_STS_IOERR;
 808                                         goto out;
 809                                 }
 810                                 ret = -EIO;
 811                                 goto out;
 812                         }
 813                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 814                                 req, bvec.bv_len);
 815                 }
 816         }
 817 out:
 818         trace_nbd_payload_received(req, handle);
 819         mutex_unlock(&cmd->lock);
 820         return ret ? ERR_PTR(ret) : cmd;
 821 }
 822
 823 static void recv_work(struct work_struct *work)
 824 {
 825         struct recv_thread_args *args = container_of(work,
 826                                                      struct recv_thread_args,
 827                                                      work);
 828         struct nbd_device *nbd = args->nbd;
 829         struct nbd_config *config = nbd->config;
 830         struct request_queue *q = nbd->disk->queue;
 831         struct nbd_sock *nsock;
 832         struct nbd_cmd *cmd;
 833         struct request *rq;
 834
 835         while (1) {
 836                 struct nbd_reply reply;
 837
 838                 if (nbd_read_reply(nbd, args->index, &reply))
 839                         break;
 840
 841                 /*
 842                  * Grab .q_usage_counter so request pool won't go away, then no
 843                  * request use-after-free is possible during nbd_handle_reply().
 844                  * If queue is frozen, there won't be any inflight requests, we
 845                  * needn't to handle the incoming garbage message.
 846                  */
 847                 if (!percpu_ref_tryget(&q->q_usage_counter)) {
 848                         dev_err(disk_to_dev(nbd->disk), "%s: no io inflight\n",
 849                                 __func__);
 850                         break;
 851                 }
 852
 853                 cmd = nbd_handle_reply(nbd, args->index, &reply);
 854                 if (IS_ERR(cmd)) {
 855                         percpu_ref_put(&q->q_usage_counter);
 856                         break;
 857                 }
 858
 859                 rq = blk_mq_rq_from_pdu(cmd);
 860                 if (likely(!blk_should_fake_timeout(rq->q))) {
 861                         bool complete;
 862
 863                         mutex_lock(&cmd->lock);
 864                         complete = __test_and_clear_bit(NBD_CMD_INFLIGHT,
 865                                                         &cmd->flags);
 866                         mutex_unlock(&cmd->lock);
 867                         if (complete)
 868                                 blk_mq_complete_request(rq);
 869                 }
 870                 percpu_ref_put(&q->q_usage_counter);
 871         }
 872
 873         nsock = config->socks[args->index];
 874         mutex_lock(&nsock->tx_lock);
 875         nbd_mark_nsock_dead(nbd, nsock, 1);
 876         mutex_unlock(&nsock->tx_lock);
 877
 878         nbd_config_put(nbd);
 879         atomic_dec(&config->recv_threads);
 880         wake_up(&config->recv_wq);
 881         kfree(args);
 882 }
 883
 884 static bool nbd_clear_req(struct request *req, void *data)
 885 {
 886         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 887
 888         /* don't abort one completed request */
 889         if (blk_mq_request_completed(req))
 890                 return true;
 891
 892         mutex_lock(&cmd->lock);
 893         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 894                 mutex_unlock(&cmd->lock);
 895                 return true;
 896         }
 897         cmd->status = BLK_STS_IOERR;
 898         mutex_unlock(&cmd->lock);
 899
 900         blk_mq_complete_request(req);
 901         return true;
 902 }
 903
 904 static void nbd_clear_que(struct nbd_device *nbd)
 905 {
 906         blk_mq_quiesce_queue(nbd->disk->queue);
 907         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 908         blk_mq_unquiesce_queue(nbd->disk->queue);
 909         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 910 }
 911
 912 static int find_fallback(struct nbd_device *nbd, int index)
 913 {
 914         struct nbd_config *config = nbd->config;
 915         int new_index = -1;
 916         struct nbd_sock *nsock = config->socks[index];
 917         int fallback = nsock->fallback_index;
 918
 919         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 920                 return new_index;
 921
 922         if (config->num_connections <= 1) {
 923                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 924                                     "Dead connection, failed to find a fallback\n");
 925                 return new_index;
 926         }
 927
 928         if (fallback >= 0 && fallback < config->num_connections &&
 929             !config->socks[fallback]->dead)
 930                 return fallback;
 931
 932         if (nsock->fallback_index < 0 ||
 933             nsock->fallback_index >= config->num_connections ||
 934             config->socks[nsock->fallback_index]->dead) {
 935                 int i;
 936                 for (i = 0; i < config->num_connections; i++) {
 937                         if (i == index)
 938                                 continue;
 939                         if (!config->socks[i]->dead) {
 940                                 new_index = i;
 941                                 break;
 942                         }
 943                 }
 944                 nsock->fallback_index = new_index;
 945                 if (new_index < 0) {
 946                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 947                                             "Dead connection, failed to find a fallback\n");
 948                         return new_index;
 949                 }
 950         }
 951         new_index = nsock->fallback_index;
 952         return new_index;
 953 }
 954
 955 static int wait_for_reconnect(struct nbd_device *nbd)
 956 {
 957         struct nbd_config *config = nbd->config;
 958         if (!config->dead_conn_timeout)
 959                 return 0;
 960
 961         if (!wait_event_timeout(config->conn_wait,
 962                                 test_bit(NBD_RT_DISCONNECTED,
 963                                          &config->runtime_flags) ||
 964                                 atomic_read(&config->live_connections) > 0,
 965                                 config->dead_conn_timeout))
 966                 return 0;
 967
 968         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
 969 }
 970
 971 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 972 {
 973         struct request *req = blk_mq_rq_from_pdu(cmd);
 974         struct nbd_device *nbd = cmd->nbd;
 975         struct nbd_config *config;
 976         struct nbd_sock *nsock;
 977         int ret;
 978
 979         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 980                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 981                                     "Socks array is empty\n");
 982                 return -EINVAL;
 983         }
 984         config = nbd->config;
 985
 986         if (index >= config->num_connections) {
 987                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 988                                     "Attempted send on invalid socket\n");
 989                 nbd_config_put(nbd);
 990                 return -EINVAL;
 991         }
 992         cmd->status = BLK_STS_OK;
 993 again:
 994         nsock = config->socks[index];
 995         mutex_lock(&nsock->tx_lock);
 996         if (nsock->dead) {
 997                 int old_index = index;
 998                 index = find_fallback(nbd, index);
 999                 mutex_unlock(&nsock->tx_lock);
1000                 if (index < 0) {
1001                         if (wait_for_reconnect(nbd)) {
1002                                 index = old_index;
1003                                 goto again;
1004                         }
1005                         /* All the sockets should already be down at this point,
1006                          * we just want to make sure that DISCONNECTED is set so
1007                          * any requests that come in that were queue'ed waiting
1008                          * for the reconnect timer don't trigger the timer again
1009                          * and instead just error out.
1010                          */
1011                         sock_shutdown(nbd);
1012                         nbd_config_put(nbd);
1013                         return -EIO;
1014                 }
1015                 goto again;
1016         }
1017
1018         /* Handle the case that we have a pending request that was partially
1019          * transmitted that _has_ to be serviced first.  We need to call requeue
1020          * here so that it gets put _after_ the request that is already on the
1021          * dispatch list.
1022          */
1023         blk_mq_start_request(req);
1024         if (unlikely(nsock->pending && nsock->pending != req)) {
1025                 nbd_requeue_cmd(cmd);
1026                 ret = 0;
1027                 goto out;
1028         }
1029         /*
1030          * Some failures are related to the link going down, so anything that
1031          * returns EAGAIN can be retried on a different socket.
1032          */
1033         ret = nbd_send_cmd(nbd, cmd, index);
1034         /*
1035          * Access to this flag is protected by cmd->lock, thus it's safe to set
1036          * the flag after nbd_send_cmd() succeed to send request to server.
1037          */
1038         if (!ret)
1039                 __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
1040         else if (ret == -EAGAIN) {
1041                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1042                                     "Request send failed, requeueing\n");
1043                 nbd_mark_nsock_dead(nbd, nsock, 1);
1044                 nbd_requeue_cmd(cmd);
1045                 ret = 0;
1046         }
1047 out:
1048         mutex_unlock(&nsock->tx_lock);
1049         nbd_config_put(nbd);
1050         return ret;
1051 }
1052
1053 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1054                         const struct blk_mq_queue_data *bd)
1055 {
1056         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1057         int ret;
1058
1059         /*
1060          * Since we look at the bio's to send the request over the network we
1061          * need to make sure the completion work doesn't mark this request done
1062          * before we are done doing our send.  This keeps us from dereferencing
1063          * freed data if we have particularly fast completions (ie we get the
1064          * completion before we exit sock_xmit on the last bvec) or in the case
1065          * that the server is misbehaving (or there was an error) before we're
1066          * done sending everything over the wire.
1067          */
1068         mutex_lock(&cmd->lock);
1069         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1070
1071         /* We can be called directly from the user space process, which means we
1072          * could possibly have signals pending so our sendmsg will fail.  In
1073          * this case we need to return that we are busy, otherwise error out as
1074          * appropriate.
1075          */
1076         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1077         if (ret < 0)
1078                 ret = BLK_STS_IOERR;
1079         else if (!ret)
1080                 ret = BLK_STS_OK;
1081         mutex_unlock(&cmd->lock);
1082
1083         return ret;
1084 }
1085
1086 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1087                                      int *err)
1088 {
1089         struct socket *sock;
1090
1091         *err = 0;
1092         sock = sockfd_lookup(fd, err);
1093         if (!sock)
1094                 return NULL;
1095
1096         if (sock->ops->shutdown == sock_no_shutdown) {
1097                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1098                 *err = -EINVAL;
1099                 sockfd_put(sock);
1100                 return NULL;
1101         }
1102
1103         return sock;
1104 }
1105
1106 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1107                           bool netlink)
1108 {
1109         struct nbd_config *config = nbd->config;
1110         struct socket *sock;
1111         struct nbd_sock **socks;
1112         struct nbd_sock *nsock;
1113         int err;
1114
1115         sock = nbd_get_socket(nbd, arg, &err);
1116         if (!sock)
1117                 return err;
1118
1119         /*
1120          * We need to make sure we don't get any errant requests while we're
1121          * reallocating the ->socks array.
1122          */
1123         blk_mq_freeze_queue(nbd->disk->queue);
1124
1125         if (!netlink && !nbd->task_setup &&
1126             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1127                 nbd->task_setup = current;
1128
1129         if (!netlink &&
1130             (nbd->task_setup != current ||
1131              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1132                 dev_err(disk_to_dev(nbd->disk),
1133                         "Device being setup by another task");
1134                 err = -EBUSY;
1135                 goto put_socket;
1136         }
1137
1138         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1139         if (!nsock) {
1140                 err = -ENOMEM;
1141                 goto put_socket;
1142         }
1143
1144         socks = krealloc(config->socks, (config->num_connections + 1) *
1145                          sizeof(struct nbd_sock *), GFP_KERNEL);
1146         if (!socks) {
1147                 kfree(nsock);
1148                 err = -ENOMEM;
1149                 goto put_socket;
1150         }
1151
1152         config->socks = socks;
1153
1154         nsock->fallback_index = -1;
1155         nsock->dead = false;
1156         mutex_init(&nsock->tx_lock);
1157         nsock->sock = sock;
1158         nsock->pending = NULL;
1159         nsock->sent = 0;
1160         nsock->cookie = 0;
1161         socks[config->num_connections++] = nsock;
1162         atomic_inc(&config->live_connections);
1163         blk_mq_unfreeze_queue(nbd->disk->queue);
1164
1165         return 0;
1166
1167 put_socket:
1168         blk_mq_unfreeze_queue(nbd->disk->queue);
1169         sockfd_put(sock);
1170         return err;
1171 }
1172
1173 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1174 {
1175         struct nbd_config *config = nbd->config;
1176         struct socket *sock, *old;
1177         struct recv_thread_args *args;
1178         int i;
1179         int err;
1180
1181         sock = nbd_get_socket(nbd, arg, &err);
1182         if (!sock)
1183                 return err;
1184
1185         args = kzalloc(sizeof(*args), GFP_KERNEL);
1186         if (!args) {
1187                 sockfd_put(sock);
1188                 return -ENOMEM;
1189         }
1190
1191         for (i = 0; i < config->num_connections; i++) {
1192                 struct nbd_sock *nsock = config->socks[i];
1193
1194                 if (!nsock->dead)
1195                         continue;
1196
1197                 mutex_lock(&nsock->tx_lock);
1198                 if (!nsock->dead) {
1199                         mutex_unlock(&nsock->tx_lock);
1200                         continue;
1201                 }
1202                 sk_set_memalloc(sock->sk);
1203                 if (nbd->tag_set.timeout)
1204                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1205                 atomic_inc(&config->recv_threads);
1206                 refcount_inc(&nbd->config_refs);
1207                 old = nsock->sock;
1208                 nsock->fallback_index = -1;
1209                 nsock->sock = sock;
1210                 nsock->dead = false;
1211                 INIT_WORK(&args->work, recv_work);
1212                 args->index = i;
1213                 args->nbd = nbd;
1214                 nsock->cookie++;
1215                 mutex_unlock(&nsock->tx_lock);
1216                 sockfd_put(old);
1217
1218                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1219
1220                 /* We take the tx_mutex in an error path in the recv_work, so we
1221                  * need to queue_work outside of the tx_mutex.
1222                  */
1223                 queue_work(nbd->recv_workq, &args->work);
1224
1225                 atomic_inc(&config->live_connections);
1226                 wake_up(&config->conn_wait);
1227                 return 0;
1228         }
1229         sockfd_put(sock);
1230         kfree(args);
1231         return -ENOSPC;
1232 }
1233
1234 static void nbd_bdev_reset(struct nbd_device *nbd)
1235 {
1236         if (disk_openers(nbd->disk) > 1)
1237                 return;
1238         set_capacity(nbd->disk, 0);
1239 }
1240
1241 static void nbd_parse_flags(struct nbd_device *nbd)
1242 {
1243         struct nbd_config *config = nbd->config;
1244         if (config->flags & NBD_FLAG_READ_ONLY)
1245                 set_disk_ro(nbd->disk, true);
1246         else
1247                 set_disk_ro(nbd->disk, false);
1248         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1249                 if (config->flags & NBD_FLAG_SEND_FUA)
1250                         blk_queue_write_cache(nbd->disk->queue, true, true);
1251                 else
1252                         blk_queue_write_cache(nbd->disk->queue, true, false);
1253         }
1254         else
1255                 blk_queue_write_cache(nbd->disk->queue, false, false);
1256 }
1257
1258 static void send_disconnects(struct nbd_device *nbd)
1259 {
1260         struct nbd_config *config = nbd->config;
1261         struct nbd_request request = {
1262                 .magic = htonl(NBD_REQUEST_MAGIC),
1263                 .type = htonl(NBD_CMD_DISC),
1264         };
1265         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1266         struct iov_iter from;
1267         int i, ret;
1268
1269         for (i = 0; i < config->num_connections; i++) {
1270                 struct nbd_sock *nsock = config->socks[i];
1271
1272                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1273                 mutex_lock(&nsock->tx_lock);
1274                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1275                 if (ret < 0)
1276                         dev_err(disk_to_dev(nbd->disk),
1277                                 "Send disconnect failed %d\n", ret);
1278                 mutex_unlock(&nsock->tx_lock);
1279         }
1280 }
1281
1282 static int nbd_disconnect(struct nbd_device *nbd)
1283 {
1284         struct nbd_config *config = nbd->config;
1285
1286         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1287         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1288         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1289         send_disconnects(nbd);
1290         return 0;
1291 }
1292
1293 static void nbd_clear_sock(struct nbd_device *nbd)
1294 {
1295         sock_shutdown(nbd);
1296         nbd_clear_que(nbd);
1297         nbd->task_setup = NULL;
1298 }
1299
1300 static void nbd_config_put(struct nbd_device *nbd)
1301 {
1302         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1303                                         &nbd->config_lock)) {
1304                 struct nbd_config *config = nbd->config;
1305                 nbd_dev_dbg_close(nbd);
1306                 invalidate_disk(nbd->disk);
1307                 if (nbd->config->bytesize)
1308                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1309                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1310                                        &config->runtime_flags))
1311                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1312                 nbd->pid = 0;
1313                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1314                                        &config->runtime_flags)) {
1315                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1316                         kfree(nbd->backend);
1317                         nbd->backend = NULL;
1318                 }
1319                 nbd_clear_sock(nbd);
1320                 if (config->num_connections) {
1321                         int i;
1322                         for (i = 0; i < config->num_connections; i++) {
1323                                 sockfd_put(config->socks[i]->sock);
1324                                 kfree(config->socks[i]);
1325                         }
1326                         kfree(config->socks);
1327                 }
1328                 kfree(nbd->config);
1329                 nbd->config = NULL;
1330
1331                 nbd->tag_set.timeout = 0;
1332                 nbd->disk->queue->limits.discard_granularity = 0;
1333                 blk_queue_max_discard_sectors(nbd->disk->queue, 0);
1334
1335                 mutex_unlock(&nbd->config_lock);
1336                 nbd_put(nbd);
1337                 module_put(THIS_MODULE);
1338         }
1339 }
1340
1341 static int nbd_start_device(struct nbd_device *nbd)
1342 {
1343         struct nbd_config *config = nbd->config;
1344         int num_connections = config->num_connections;
1345         int error = 0, i;
1346
1347         if (nbd->pid)
1348                 return -EBUSY;
1349         if (!config->socks)
1350                 return -EINVAL;
1351         if (num_connections > 1 &&
1352             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1353                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1354                 return -EINVAL;
1355         }
1356
1357         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1358         nbd->pid = task_pid_nr(current);
1359
1360         nbd_parse_flags(nbd);
1361
1362         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1363         if (error) {
1364                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1365                 return error;
1366         }
1367         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1368
1369         nbd_dev_dbg_init(nbd);
1370         for (i = 0; i < num_connections; i++) {
1371                 struct recv_thread_args *args;
1372
1373                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1374                 if (!args) {
1375                         sock_shutdown(nbd);
1376                         /*
1377                          * If num_connections is m (2 < m),
1378                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1379                          * But NO.(n + 1) failed. We still have n recv threads.
1380                          * So, add flush_workqueue here to prevent recv threads
1381                          * dropping the last config_refs and trying to destroy
1382                          * the workqueue from inside the workqueue.
1383                          */
1384                         if (i)
1385                                 flush_workqueue(nbd->recv_workq);
1386                         return -ENOMEM;
1387                 }
1388                 sk_set_memalloc(config->socks[i]->sock->sk);
1389                 if (nbd->tag_set.timeout)
1390                         config->socks[i]->sock->sk->sk_sndtimeo =
1391                                 nbd->tag_set.timeout;
1392                 atomic_inc(&config->recv_threads);
1393                 refcount_inc(&nbd->config_refs);
1394                 INIT_WORK(&args->work, recv_work);
1395                 args->nbd = nbd;
1396                 args->index = i;
1397                 queue_work(nbd->recv_workq, &args->work);
1398         }
1399         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1400 }
1401
1402 static int nbd_start_device_ioctl(struct nbd_device *nbd)
1403 {
1404         struct nbd_config *config = nbd->config;
1405         int ret;
1406
1407         ret = nbd_start_device(nbd);
1408         if (ret)
1409                 return ret;
1410
1411         if (max_part)
1412                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1413         mutex_unlock(&nbd->config_lock);
1414         ret = wait_event_interruptible(config->recv_wq,
1415                                          atomic_read(&config->recv_threads) == 0);
1416         if (ret)
1417                 sock_shutdown(nbd);
1418         flush_workqueue(nbd->recv_workq);
1419
1420         mutex_lock(&nbd->config_lock);
1421         nbd_bdev_reset(nbd);
1422         /* user requested, ignore socket errors */
1423         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1424                 ret = 0;
1425         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1426                 ret = -ETIMEDOUT;
1427         return ret;
1428 }
1429
1430 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1431                                  struct block_device *bdev)
1432 {
1433         nbd_clear_sock(nbd);
1434         __invalidate_device(bdev, true);
1435         nbd_bdev_reset(nbd);
1436         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1437                                &nbd->config->runtime_flags))
1438                 nbd_config_put(nbd);
1439 }
1440
1441 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1442 {
1443         nbd->tag_set.timeout = timeout * HZ;
1444         if (timeout)
1445                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1446         else
1447                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1448 }
1449
1450 /* Must be called with config_lock held */
1451 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1452                        unsigned int cmd, unsigned long arg)
1453 {
1454         struct nbd_config *config = nbd->config;
1455         loff_t bytesize;
1456
1457         switch (cmd) {
1458         case NBD_DISCONNECT:
1459                 return nbd_disconnect(nbd);
1460         case NBD_CLEAR_SOCK:
1461                 nbd_clear_sock_ioctl(nbd, bdev);
1462                 return 0;
1463         case NBD_SET_SOCK:
1464                 return nbd_add_socket(nbd, arg, false);
1465         case NBD_SET_BLKSIZE:
1466                 return nbd_set_size(nbd, config->bytesize, arg);
1467         case NBD_SET_SIZE:
1468                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1469         case NBD_SET_SIZE_BLOCKS:
1470                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1471                         return -EINVAL;
1472                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1473         case NBD_SET_TIMEOUT:
1474                 nbd_set_cmd_timeout(nbd, arg);
1475                 return 0;
1476
1477         case NBD_SET_FLAGS:
1478                 config->flags = arg;
1479                 return 0;
1480         case NBD_DO_IT:
1481                 return nbd_start_device_ioctl(nbd);
1482         case NBD_CLEAR_QUE:
1483                 /*
1484                  * This is for compatibility only.  The queue is always cleared
1485                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1486                  */
1487                 return 0;
1488         case NBD_PRINT_DEBUG:
1489                 /*
1490                  * For compatibility only, we no longer keep a list of
1491                  * outstanding requests.
1492                  */
1493                 return 0;
1494         }
1495         return -ENOTTY;
1496 }
1497
1498 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1499                      unsigned int cmd, unsigned long arg)
1500 {
1501         struct nbd_device *nbd = bdev->bd_disk->private_data;
1502         struct nbd_config *config = nbd->config;
1503         int error = -EINVAL;
1504
1505         if (!capable(CAP_SYS_ADMIN))
1506                 return -EPERM;
1507
1508         /* The block layer will pass back some non-nbd ioctls in case we have
1509          * special handling for them, but we don't so just return an error.
1510          */
1511         if (_IOC_TYPE(cmd) != 0xab)
1512                 return -EINVAL;
1513
1514         mutex_lock(&nbd->config_lock);
1515
1516         /* Don't allow ioctl operations on a nbd device that was created with
1517          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1518          */
1519         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1520             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1521                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1522         else
1523                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1524         mutex_unlock(&nbd->config_lock);
1525         return error;
1526 }
1527
1528 static struct nbd_config *nbd_alloc_config(void)
1529 {
1530         struct nbd_config *config;
1531
1532         if (!try_module_get(THIS_MODULE))
1533                 return ERR_PTR(-ENODEV);
1534
1535         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1536         if (!config) {
1537                 module_put(THIS_MODULE);
1538                 return ERR_PTR(-ENOMEM);
1539         }
1540
1541         atomic_set(&config->recv_threads, 0);
1542         init_waitqueue_head(&config->recv_wq);
1543         init_waitqueue_head(&config->conn_wait);
1544         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1545         atomic_set(&config->live_connections, 0);
1546         return config;
1547 }
1548
1549 static int nbd_open(struct block_device *bdev, fmode_t mode)
1550 {
1551         struct nbd_device *nbd;
1552         int ret = 0;
1553
1554         mutex_lock(&nbd_index_mutex);
1555         nbd = bdev->bd_disk->private_data;
1556         if (!nbd) {
1557                 ret = -ENXIO;
1558                 goto out;
1559         }
1560         if (!refcount_inc_not_zero(&nbd->refs)) {
1561                 ret = -ENXIO;
1562                 goto out;
1563         }
1564         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1565                 struct nbd_config *config;
1566
1567                 mutex_lock(&nbd->config_lock);
1568                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1569                         mutex_unlock(&nbd->config_lock);
1570                         goto out;
1571                 }
1572                 config = nbd_alloc_config();
1573                 if (IS_ERR(config)) {
1574                         ret = PTR_ERR(config);
1575                         mutex_unlock(&nbd->config_lock);
1576                         goto out;
1577                 }
1578                 nbd->config = config;
1579                 refcount_set(&nbd->config_refs, 1);
1580                 refcount_inc(&nbd->refs);
1581                 mutex_unlock(&nbd->config_lock);
1582                 if (max_part)
1583                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1584         } else if (nbd_disconnected(nbd->config)) {
1585                 if (max_part)
1586                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1587         }
1588 out:
1589         mutex_unlock(&nbd_index_mutex);
1590         return ret;
1591 }
1592
1593 static void nbd_release(struct gendisk *disk, fmode_t mode)
1594 {
1595         struct nbd_device *nbd = disk->private_data;
1596
1597         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1598                         disk_openers(disk) == 0)
1599                 nbd_disconnect_and_put(nbd);
1600
1601         nbd_config_put(nbd);
1602         nbd_put(nbd);
1603 }
1604
1605 static const struct block_device_operations nbd_fops =
1606 {
1607         .owner =        THIS_MODULE,
1608         .open =         nbd_open,
1609         .release =      nbd_release,
1610         .ioctl =        nbd_ioctl,
1611         .compat_ioctl = nbd_ioctl,
1612 };
1613
1614 #if IS_ENABLED(CONFIG_DEBUG_FS)
1615
1616 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1617 {
1618         struct nbd_device *nbd = s->private;
1619
1620         if (nbd->pid)
1621                 seq_printf(s, "recv: %d\n", nbd->pid);
1622
1623         return 0;
1624 }
1625
1626 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1627
1628 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1629 {
1630         struct nbd_device *nbd = s->private;
1631         u32 flags = nbd->config->flags;
1632
1633         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1634
1635         seq_puts(s, "Known flags:\n");
1636
1637         if (flags & NBD_FLAG_HAS_FLAGS)
1638                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1639         if (flags & NBD_FLAG_READ_ONLY)
1640                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1641         if (flags & NBD_FLAG_SEND_FLUSH)
1642                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1643         if (flags & NBD_FLAG_SEND_FUA)
1644                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1645         if (flags & NBD_FLAG_SEND_TRIM)
1646                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1647
1648         return 0;
1649 }
1650
1651 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1652
1653 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1654 {
1655         struct dentry *dir;
1656         struct nbd_config *config = nbd->config;
1657
1658         if (!nbd_dbg_dir)
1659                 return -EIO;
1660
1661         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1662         if (!dir) {
1663                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1664                         nbd_name(nbd));
1665                 return -EIO;
1666         }
1667         config->dbg_dir = dir;
1668
1669         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1670         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1671         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1672         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1673         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1674
1675         return 0;
1676 }
1677
1678 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1679 {
1680         debugfs_remove_recursive(nbd->config->dbg_dir);
1681 }
1682
1683 static int nbd_dbg_init(void)
1684 {
1685         struct dentry *dbg_dir;
1686
1687         dbg_dir = debugfs_create_dir("nbd", NULL);
1688         if (!dbg_dir)
1689                 return -EIO;
1690
1691         nbd_dbg_dir = dbg_dir;
1692
1693         return 0;
1694 }
1695
1696 static void nbd_dbg_close(void)
1697 {
1698         debugfs_remove_recursive(nbd_dbg_dir);
1699 }
1700
1701 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1702
1703 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1704 {
1705         return 0;
1706 }
1707
1708 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1709 {
1710 }
1711
1712 static int nbd_dbg_init(void)
1713 {
1714         return 0;
1715 }
1716
1717 static void nbd_dbg_close(void)
1718 {
1719 }
1720
1721 #endif
1722
1723 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1724                             unsigned int hctx_idx, unsigned int numa_node)
1725 {
1726         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1727         cmd->nbd = set->driver_data;
1728         cmd->flags = 0;
1729         mutex_init(&cmd->lock);
1730         return 0;
1731 }
1732
1733 static const struct blk_mq_ops nbd_mq_ops = {
1734         .queue_rq       = nbd_queue_rq,
1735         .complete       = nbd_complete_rq,
1736         .init_request   = nbd_init_request,
1737         .timeout        = nbd_xmit_timeout,
1738 };
1739
1740 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1741 {
1742         struct nbd_device *nbd;
1743         struct gendisk *disk;
1744         int err = -ENOMEM;
1745
1746         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1747         if (!nbd)
1748                 goto out;
1749
1750         nbd->tag_set.ops = &nbd_mq_ops;
1751         nbd->tag_set.nr_hw_queues = 1;
1752         nbd->tag_set.queue_depth = 128;
1753         nbd->tag_set.numa_node = NUMA_NO_NODE;
1754         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1755         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1756                 BLK_MQ_F_BLOCKING;
1757         nbd->tag_set.driver_data = nbd;
1758         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1759         nbd->backend = NULL;
1760
1761         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1762         if (err)
1763                 goto out_free_nbd;
1764
1765         mutex_lock(&nbd_index_mutex);
1766         if (index >= 0) {
1767                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1768                                 GFP_KERNEL);
1769                 if (err == -ENOSPC)
1770                         err = -EEXIST;
1771         } else {
1772                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1773                 if (err >= 0)
1774                         index = err;
1775         }
1776         nbd->index = index;
1777         mutex_unlock(&nbd_index_mutex);
1778         if (err < 0)
1779                 goto out_free_tags;
1780
1781         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1782         if (IS_ERR(disk)) {
1783                 err = PTR_ERR(disk);
1784                 goto out_free_idr;
1785         }
1786         nbd->disk = disk;
1787
1788         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1789                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1790                                           WQ_UNBOUND, 0, nbd->index);
1791         if (!nbd->recv_workq) {
1792                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1793                 err = -ENOMEM;
1794                 goto out_err_disk;
1795         }
1796
1797         /*
1798          * Tell the block layer that we are not a rotational device
1799          */
1800         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1801         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1802         disk->queue->limits.discard_granularity = 0;
1803         blk_queue_max_discard_sectors(disk->queue, 0);
1804         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1805         blk_queue_max_segments(disk->queue, USHRT_MAX);
1806         blk_queue_max_hw_sectors(disk->queue, 65536);
1807         disk->queue->limits.max_sectors = 256;
1808
1809         mutex_init(&nbd->config_lock);
1810         refcount_set(&nbd->config_refs, 0);
1811         /*
1812          * Start out with a zero references to keep other threads from using
1813          * this device until it is fully initialized.
1814          */
1815         refcount_set(&nbd->refs, 0);
1816         INIT_LIST_HEAD(&nbd->list);
1817         disk->major = NBD_MAJOR;
1818         disk->first_minor = index << part_shift;
1819         disk->minors = 1 << part_shift;
1820         disk->fops = &nbd_fops;
1821         disk->private_data = nbd;
1822         sprintf(disk->disk_name, "nbd%d", index);
1823         err = add_disk(disk);
1824         if (err)
1825                 goto out_free_work;
1826
1827         /*
1828          * Now publish the device.
1829          */
1830         refcount_set(&nbd->refs, refs);
1831         nbd_total_devices++;
1832         return nbd;
1833
1834 out_free_work:
1835         destroy_workqueue(nbd->recv_workq);
1836 out_err_disk:
1837         put_disk(disk);
1838 out_free_idr:
1839         mutex_lock(&nbd_index_mutex);
1840         idr_remove(&nbd_index_idr, index);
1841         mutex_unlock(&nbd_index_mutex);
1842 out_free_tags:
1843         blk_mq_free_tag_set(&nbd->tag_set);
1844 out_free_nbd:
1845         kfree(nbd);
1846 out:
1847         return ERR_PTR(err);
1848 }
1849
1850 static struct nbd_device *nbd_find_get_unused(void)
1851 {
1852         struct nbd_device *nbd;
1853         int id;
1854
1855         lockdep_assert_held(&nbd_index_mutex);
1856
1857         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1858                 if (refcount_read(&nbd->config_refs) ||
1859                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1860                         continue;
1861                 if (refcount_inc_not_zero(&nbd->refs))
1862                         return nbd;
1863         }
1864
1865         return NULL;
1866 }
1867
1868 /* Netlink interface. */
1869 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1870         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1871         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1872         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1873         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1874         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1875         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1876         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1877         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1878         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1879         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1880 };
1881
1882 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1883         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1884 };
1885
1886 /* We don't use this right now since we don't parse the incoming list, but we
1887  * still want it here so userspace knows what to expect.
1888  */
1889 static const struct nla_policy __attribute__((unused))
1890 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1891         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1892         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1893 };
1894
1895 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1896 {
1897         struct nbd_config *config = nbd->config;
1898         u64 bsize = nbd_blksize(config);
1899         u64 bytes = config->bytesize;
1900
1901         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1902                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1903
1904         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1905                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1906
1907         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1908                 return nbd_set_size(nbd, bytes, bsize);
1909         return 0;
1910 }
1911
1912 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1913 {
1914         struct nbd_device *nbd;
1915         struct nbd_config *config;
1916         int index = -1;
1917         int ret;
1918         bool put_dev = false;
1919
1920         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1921                 return -EPERM;
1922
1923         if (info->attrs[NBD_ATTR_INDEX]) {
1924                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1925
1926                 /*
1927                  * Too big first_minor can cause duplicate creation of
1928                  * sysfs files/links, since index << part_shift might overflow, or
1929                  * MKDEV() expect that the max bits of first_minor is 20.
1930                  */
1931                 if (index < 0 || index > MINORMASK >> part_shift) {
1932                         pr_err("illegal input index %d\n", index);
1933                         return -EINVAL;
1934                 }
1935         }
1936         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1937                 pr_err("must specify at least one socket\n");
1938                 return -EINVAL;
1939         }
1940         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1941                 pr_err("must specify a size in bytes for the device\n");
1942                 return -EINVAL;
1943         }
1944 again:
1945         mutex_lock(&nbd_index_mutex);
1946         if (index == -1) {
1947                 nbd = nbd_find_get_unused();
1948         } else {
1949                 nbd = idr_find(&nbd_index_idr, index);
1950                 if (nbd) {
1951                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1952                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1953                             !refcount_inc_not_zero(&nbd->refs)) {
1954                                 mutex_unlock(&nbd_index_mutex);
1955                                 pr_err("device at index %d is going down\n",
1956                                         index);
1957                                 return -EINVAL;
1958                         }
1959                 }
1960         }
1961         mutex_unlock(&nbd_index_mutex);
1962
1963         if (!nbd) {
1964                 nbd = nbd_dev_add(index, 2);
1965                 if (IS_ERR(nbd)) {
1966                         pr_err("failed to add new device\n");
1967                         return PTR_ERR(nbd);
1968                 }
1969         }
1970
1971         mutex_lock(&nbd->config_lock);
1972         if (refcount_read(&nbd->config_refs)) {
1973                 mutex_unlock(&nbd->config_lock);
1974                 nbd_put(nbd);
1975                 if (index == -1)
1976                         goto again;
1977                 pr_err("nbd%d already in use\n", index);
1978                 return -EBUSY;
1979         }
1980         if (WARN_ON(nbd->config)) {
1981                 mutex_unlock(&nbd->config_lock);
1982                 nbd_put(nbd);
1983                 return -EINVAL;
1984         }
1985         config = nbd_alloc_config();
1986         if (IS_ERR(config)) {
1987                 mutex_unlock(&nbd->config_lock);
1988                 nbd_put(nbd);
1989                 pr_err("couldn't allocate config\n");
1990                 return PTR_ERR(config);
1991         }
1992         nbd->config = config;
1993         refcount_set(&nbd->config_refs, 1);
1994         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1995
1996         ret = nbd_genl_size_set(info, nbd);
1997         if (ret)
1998                 goto out;
1999
2000         if (info->attrs[NBD_ATTR_TIMEOUT])
2001                 nbd_set_cmd_timeout(nbd,
2002                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2003         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2004                 config->dead_conn_timeout =
2005                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2006                 config->dead_conn_timeout *= HZ;
2007         }
2008         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
2009                 config->flags =
2010                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
2011         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2012                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2013                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2014                         /*
2015                          * We have 1 ref to keep the device around, and then 1
2016                          * ref for our current operation here, which will be
2017                          * inherited by the config.  If we already have
2018                          * DESTROY_ON_DISCONNECT set then we know we don't have
2019                          * that extra ref already held so we don't need the
2020                          * put_dev.
2021                          */
2022                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2023                                               &nbd->flags))
2024                                 put_dev = true;
2025                 } else {
2026                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2027                                                &nbd->flags))
2028                                 refcount_inc(&nbd->refs);
2029                 }
2030                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2031                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2032                                 &config->runtime_flags);
2033                 }
2034         }
2035
2036         if (info->attrs[NBD_ATTR_SOCKETS]) {
2037                 struct nlattr *attr;
2038                 int rem, fd;
2039
2040                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2041                                     rem) {
2042                         struct nlattr *socks[NBD_SOCK_MAX+1];
2043
2044                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2045                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2046                                 ret = -EINVAL;
2047                                 goto out;
2048                         }
2049                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2050                                                           attr,
2051                                                           nbd_sock_policy,
2052                                                           info->extack);
2053                         if (ret != 0) {
2054                                 pr_err("error processing sock list\n");
2055                                 ret = -EINVAL;
2056                                 goto out;
2057                         }
2058                         if (!socks[NBD_SOCK_FD])
2059                                 continue;
2060                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2061                         ret = nbd_add_socket(nbd, fd, true);
2062                         if (ret)
2063                                 goto out;
2064                 }
2065         }
2066         ret = nbd_start_device(nbd);
2067         if (ret)
2068                 goto out;
2069         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2070                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2071                                           GFP_KERNEL);
2072                 if (!nbd->backend) {
2073                         ret = -ENOMEM;
2074                         goto out;
2075                 }
2076         }
2077         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2078         if (ret) {
2079                 dev_err(disk_to_dev(nbd->disk),
2080                         "device_create_file failed for backend!\n");
2081                 goto out;
2082         }
2083         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2084 out:
2085         mutex_unlock(&nbd->config_lock);
2086         if (!ret) {
2087                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2088                 refcount_inc(&nbd->config_refs);
2089                 nbd_connect_reply(info, nbd->index);
2090         }
2091         nbd_config_put(nbd);
2092         if (put_dev)
2093                 nbd_put(nbd);
2094         return ret;
2095 }
2096
2097 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2098 {
2099         mutex_lock(&nbd->config_lock);
2100         nbd_disconnect(nbd);
2101         sock_shutdown(nbd);
2102         wake_up(&nbd->config->conn_wait);
2103         /*
2104          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2105          * to cancel the inflight I/Os.
2106          */
2107         flush_workqueue(nbd->recv_workq);
2108         nbd_clear_que(nbd);
2109         nbd->task_setup = NULL;
2110         mutex_unlock(&nbd->config_lock);
2111
2112         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2113                                &nbd->config->runtime_flags))
2114                 nbd_config_put(nbd);
2115 }
2116
2117 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2118 {
2119         struct nbd_device *nbd;
2120         int index;
2121
2122         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2123                 return -EPERM;
2124
2125         if (!info->attrs[NBD_ATTR_INDEX]) {
2126                 pr_err("must specify an index to disconnect\n");
2127                 return -EINVAL;
2128         }
2129         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2130         mutex_lock(&nbd_index_mutex);
2131         nbd = idr_find(&nbd_index_idr, index);
2132         if (!nbd) {
2133                 mutex_unlock(&nbd_index_mutex);
2134                 pr_err("couldn't find device at index %d\n", index);
2135                 return -EINVAL;
2136         }
2137         if (!refcount_inc_not_zero(&nbd->refs)) {
2138                 mutex_unlock(&nbd_index_mutex);
2139                 pr_err("device at index %d is going down\n", index);
2140                 return -EINVAL;
2141         }
2142         mutex_unlock(&nbd_index_mutex);
2143         if (!refcount_inc_not_zero(&nbd->config_refs))
2144                 goto put_nbd;
2145         nbd_disconnect_and_put(nbd);
2146         nbd_config_put(nbd);
2147 put_nbd:
2148         nbd_put(nbd);
2149         return 0;
2150 }
2151
2152 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2153 {
2154         struct nbd_device *nbd = NULL;
2155         struct nbd_config *config;
2156         int index;
2157         int ret = 0;
2158         bool put_dev = false;
2159
2160         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2161                 return -EPERM;
2162
2163         if (!info->attrs[NBD_ATTR_INDEX]) {
2164                 pr_err("must specify a device to reconfigure\n");
2165                 return -EINVAL;
2166         }
2167         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2168         mutex_lock(&nbd_index_mutex);
2169         nbd = idr_find(&nbd_index_idr, index);
2170         if (!nbd) {
2171                 mutex_unlock(&nbd_index_mutex);
2172                 pr_err("couldn't find a device at index %d\n", index);
2173                 return -EINVAL;
2174         }
2175         if (nbd->backend) {
2176                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2177                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2178                                        nbd->backend)) {
2179                                 mutex_unlock(&nbd_index_mutex);
2180                                 dev_err(nbd_to_dev(nbd),
2181                                         "backend image doesn't match with %s\n",
2182                                         nbd->backend);
2183                                 return -EINVAL;
2184                         }
2185                 } else {
2186                         mutex_unlock(&nbd_index_mutex);
2187                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2188                         return -EINVAL;
2189                 }
2190         }
2191         if (!refcount_inc_not_zero(&nbd->refs)) {
2192                 mutex_unlock(&nbd_index_mutex);
2193                 pr_err("device at index %d is going down\n", index);
2194                 return -EINVAL;
2195         }
2196         mutex_unlock(&nbd_index_mutex);
2197
2198         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2199                 dev_err(nbd_to_dev(nbd),
2200                         "not configured, cannot reconfigure\n");
2201                 nbd_put(nbd);
2202                 return -EINVAL;
2203         }
2204
2205         mutex_lock(&nbd->config_lock);
2206         config = nbd->config;
2207         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2208             !nbd->pid) {
2209                 dev_err(nbd_to_dev(nbd),
2210                         "not configured, cannot reconfigure\n");
2211                 ret = -EINVAL;
2212                 goto out;
2213         }
2214
2215         ret = nbd_genl_size_set(info, nbd);
2216         if (ret)
2217                 goto out;
2218
2219         if (info->attrs[NBD_ATTR_TIMEOUT])
2220                 nbd_set_cmd_timeout(nbd,
2221                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2222         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2223                 config->dead_conn_timeout =
2224                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2225                 config->dead_conn_timeout *= HZ;
2226         }
2227         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2228                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2229                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2230                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2231                                               &nbd->flags))
2232                                 put_dev = true;
2233                 } else {
2234                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2235                                                &nbd->flags))
2236                                 refcount_inc(&nbd->refs);
2237                 }
2238
2239                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2240                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2241                                         &config->runtime_flags);
2242                 } else {
2243                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2244                                         &config->runtime_flags);
2245                 }
2246         }
2247
2248         if (info->attrs[NBD_ATTR_SOCKETS]) {
2249                 struct nlattr *attr;
2250                 int rem, fd;
2251
2252                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2253                                     rem) {
2254                         struct nlattr *socks[NBD_SOCK_MAX+1];
2255
2256                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2257                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2258                                 ret = -EINVAL;
2259                                 goto out;
2260                         }
2261                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2262                                                           attr,
2263                                                           nbd_sock_policy,
2264                                                           info->extack);
2265                         if (ret != 0) {
2266                                 pr_err("error processing sock list\n");
2267                                 ret = -EINVAL;
2268                                 goto out;
2269                         }
2270                         if (!socks[NBD_SOCK_FD])
2271                                 continue;
2272                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2273                         ret = nbd_reconnect_socket(nbd, fd);
2274                         if (ret) {
2275                                 if (ret == -ENOSPC)
2276                                         ret = 0;
2277                                 goto out;
2278                         }
2279                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2280                 }
2281         }
2282 out:
2283         mutex_unlock(&nbd->config_lock);
2284         nbd_config_put(nbd);
2285         nbd_put(nbd);
2286         if (put_dev)
2287                 nbd_put(nbd);
2288         return ret;
2289 }
2290
2291 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2292         {
2293                 .cmd    = NBD_CMD_CONNECT,
2294                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2295                 .doit   = nbd_genl_connect,
2296         },
2297         {
2298                 .cmd    = NBD_CMD_DISCONNECT,
2299                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2300                 .doit   = nbd_genl_disconnect,
2301         },
2302         {
2303                 .cmd    = NBD_CMD_RECONFIGURE,
2304                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2305                 .doit   = nbd_genl_reconfigure,
2306         },
2307         {
2308                 .cmd    = NBD_CMD_STATUS,
2309                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2310                 .doit   = nbd_genl_status,
2311         },
2312 };
2313
2314 static const struct genl_multicast_group nbd_mcast_grps[] = {
2315         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2316 };
2317
2318 static struct genl_family nbd_genl_family __ro_after_init = {
2319         .hdrsize        = 0,
2320         .name           = NBD_GENL_FAMILY_NAME,
2321         .version        = NBD_GENL_VERSION,
2322         .module         = THIS_MODULE,
2323         .small_ops      = nbd_connect_genl_ops,
2324         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2325         .maxattr        = NBD_ATTR_MAX,
2326         .policy = nbd_attr_policy,
2327         .mcgrps         = nbd_mcast_grps,
2328         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2329 };
2330
2331 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2332 {
2333         struct nlattr *dev_opt;
2334         u8 connected = 0;
2335         int ret;
2336
2337         /* This is a little racey, but for status it's ok.  The
2338          * reason we don't take a ref here is because we can't
2339          * take a ref in the index == -1 case as we would need
2340          * to put under the nbd_index_mutex, which could
2341          * deadlock if we are configured to remove ourselves
2342          * once we're disconnected.
2343          */
2344         if (refcount_read(&nbd->config_refs))
2345                 connected = 1;
2346         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2347         if (!dev_opt)
2348                 return -EMSGSIZE;
2349         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2350         if (ret)
2351                 return -EMSGSIZE;
2352         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2353                          connected);
2354         if (ret)
2355                 return -EMSGSIZE;
2356         nla_nest_end(reply, dev_opt);
2357         return 0;
2358 }
2359
2360 static int status_cb(int id, void *ptr, void *data)
2361 {
2362         struct nbd_device *nbd = ptr;
2363         return populate_nbd_status(nbd, (struct sk_buff *)data);
2364 }
2365
2366 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2367 {
2368         struct nlattr *dev_list;
2369         struct sk_buff *reply;
2370         void *reply_head;
2371         size_t msg_size;
2372         int index = -1;
2373         int ret = -ENOMEM;
2374
2375         if (info->attrs[NBD_ATTR_INDEX])
2376                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2377
2378         mutex_lock(&nbd_index_mutex);
2379
2380         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2381                                   nla_attr_size(sizeof(u8)));
2382         msg_size *= (index == -1) ? nbd_total_devices : 1;
2383
2384         reply = genlmsg_new(msg_size, GFP_KERNEL);
2385         if (!reply)
2386                 goto out;
2387         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2388                                        NBD_CMD_STATUS);
2389         if (!reply_head) {
2390                 nlmsg_free(reply);
2391                 goto out;
2392         }
2393
2394         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2395         if (index == -1) {
2396                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2397                 if (ret) {
2398                         nlmsg_free(reply);
2399                         goto out;
2400                 }
2401         } else {
2402                 struct nbd_device *nbd;
2403                 nbd = idr_find(&nbd_index_idr, index);
2404                 if (nbd) {
2405                         ret = populate_nbd_status(nbd, reply);
2406                         if (ret) {
2407                                 nlmsg_free(reply);
2408                                 goto out;
2409                         }
2410                 }
2411         }
2412         nla_nest_end(reply, dev_list);
2413         genlmsg_end(reply, reply_head);
2414         ret = genlmsg_reply(reply, info);
2415 out:
2416         mutex_unlock(&nbd_index_mutex);
2417         return ret;
2418 }
2419
2420 static void nbd_connect_reply(struct genl_info *info, int index)
2421 {
2422         struct sk_buff *skb;
2423         void *msg_head;
2424         int ret;
2425
2426         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2427         if (!skb)
2428                 return;
2429         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2430                                      NBD_CMD_CONNECT);
2431         if (!msg_head) {
2432                 nlmsg_free(skb);
2433                 return;
2434         }
2435         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2436         if (ret) {
2437                 nlmsg_free(skb);
2438                 return;
2439         }
2440         genlmsg_end(skb, msg_head);
2441         genlmsg_reply(skb, info);
2442 }
2443
2444 static void nbd_mcast_index(int index)
2445 {
2446         struct sk_buff *skb;
2447         void *msg_head;
2448         int ret;
2449
2450         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2451         if (!skb)
2452                 return;
2453         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2454                                      NBD_CMD_LINK_DEAD);
2455         if (!msg_head) {
2456                 nlmsg_free(skb);
2457                 return;
2458         }
2459         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2460         if (ret) {
2461                 nlmsg_free(skb);
2462                 return;
2463         }
2464         genlmsg_end(skb, msg_head);
2465         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2466 }
2467
2468 static void nbd_dead_link_work(struct work_struct *work)
2469 {
2470         struct link_dead_args *args = container_of(work, struct link_dead_args,
2471                                                    work);
2472         nbd_mcast_index(args->index);
2473         kfree(args);
2474 }
2475
2476 static int __init nbd_init(void)
2477 {
2478         int i;
2479
2480         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2481
2482         if (max_part < 0) {
2483                 pr_err("max_part must be >= 0\n");
2484                 return -EINVAL;
2485         }
2486
2487         part_shift = 0;
2488         if (max_part > 0) {
2489                 part_shift = fls(max_part);
2490
2491                 /*
2492                  * Adjust max_part according to part_shift as it is exported
2493                  * to user space so that user can know the max number of
2494                  * partition kernel should be able to manage.
2495                  *
2496                  * Note that -1 is required because partition 0 is reserved
2497                  * for the whole disk.
2498                  */
2499                 max_part = (1UL << part_shift) - 1;
2500         }
2501
2502         if ((1UL << part_shift) > DISK_MAX_PARTS)
2503                 return -EINVAL;
2504
2505         if (nbds_max > 1UL << (MINORBITS - part_shift))
2506                 return -EINVAL;
2507
2508         if (register_blkdev(NBD_MAJOR, "nbd"))
2509                 return -EIO;
2510
2511         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2512         if (!nbd_del_wq) {
2513                 unregister_blkdev(NBD_MAJOR, "nbd");
2514                 return -ENOMEM;
2515         }
2516
2517         if (genl_register_family(&nbd_genl_family)) {
2518                 destroy_workqueue(nbd_del_wq);
2519                 unregister_blkdev(NBD_MAJOR, "nbd");
2520                 return -EINVAL;
2521         }
2522         nbd_dbg_init();
2523
2524         for (i = 0; i < nbds_max; i++)
2525                 nbd_dev_add(i, 1);
2526         return 0;
2527 }
2528
2529 static int nbd_exit_cb(int id, void *ptr, void *data)
2530 {
2531         struct list_head *list = (struct list_head *)data;
2532         struct nbd_device *nbd = ptr;
2533
2534         /* Skip nbd that is being removed asynchronously */
2535         if (refcount_read(&nbd->refs))
2536                 list_add_tail(&nbd->list, list);
2537
2538         return 0;
2539 }
2540
2541 static void __exit nbd_cleanup(void)
2542 {
2543         struct nbd_device *nbd;
2544         LIST_HEAD(del_list);
2545
2546         /*
2547          * Unregister netlink interface prior to waiting
2548          * for the completion of netlink commands.
2549          */
2550         genl_unregister_family(&nbd_genl_family);
2551
2552         nbd_dbg_close();
2553
2554         mutex_lock(&nbd_index_mutex);
2555         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2556         mutex_unlock(&nbd_index_mutex);
2557
2558         while (!list_empty(&del_list)) {
2559                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2560                 list_del_init(&nbd->list);
2561                 if (refcount_read(&nbd->config_refs))
2562                         pr_err("possibly leaking nbd_config (ref %d)\n",
2563                                         refcount_read(&nbd->config_refs));
2564                 if (refcount_read(&nbd->refs) != 1)
2565                         pr_err("possibly leaking a device\n");
2566                 nbd_put(nbd);
2567         }
2568
2569         /* Also wait for nbd_dev_remove_work() completes */
2570         destroy_workqueue(nbd_del_wq);
2571
2572         idr_destroy(&nbd_index_idr);
2573         unregister_blkdev(NBD_MAJOR, "nbd");
2574 }
2575
2576 module_init(nbd_init);
2577 module_exit(nbd_cleanup);
2578
2579 MODULE_DESCRIPTION("Network Block Device");
2580 MODULE_LICENSE("GPL");
2581
2582 module_param(nbds_max, int, 0444);
2583 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2584 module_param(max_part, int, 0444);
2585 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");