drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #include <linux/major.h>
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/module.h>
  18 #include <linux/init.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/completion.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/nbd.h>
  49
  50 static DEFINE_IDR(nbd_index_idr);
  51 static DEFINE_MUTEX(nbd_index_mutex);
  52 static int nbd_total_devices = 0;
  53
  54 struct nbd_sock {
  55         struct socket *sock;
  56         struct mutex tx_lock;
  57         struct request *pending;
  58         int sent;
  59         bool dead;
  60         int fallback_index;
  61         int cookie;
  62 };
  63
  64 struct recv_thread_args {
  65         struct work_struct work;
  66         struct nbd_device *nbd;
  67         int index;
  68 };
  69
  70 struct link_dead_args {
  71         struct work_struct work;
  72         int index;
  73 };
  74
  75 #define NBD_RT_TIMEDOUT                 0
  76 #define NBD_RT_DISCONNECT_REQUESTED     1
  77 #define NBD_RT_DISCONNECTED             2
  78 #define NBD_RT_HAS_PID_FILE             3
  79 #define NBD_RT_HAS_CONFIG_REF           4
  80 #define NBD_RT_BOUND                    5
  81 #define NBD_RT_DESTROY_ON_DISCONNECT    6
  82 #define NBD_RT_DISCONNECT_ON_CLOSE      7
  83
  84 #define NBD_DESTROY_ON_DISCONNECT       0
  85 #define NBD_DISCONNECT_REQUESTED        1
  86
  87 struct nbd_config {
  88         u32 flags;
  89         unsigned long runtime_flags;
  90         u64 dead_conn_timeout;
  91
  92         struct nbd_sock **socks;
  93         int num_connections;
  94         atomic_t live_connections;
  95         wait_queue_head_t conn_wait;
  96
  97         atomic_t recv_threads;
  98         wait_queue_head_t recv_wq;
  99         loff_t blksize;
 100         loff_t bytesize;
 101 #if IS_ENABLED(CONFIG_DEBUG_FS)
 102         struct dentry *dbg_dir;
 103 #endif
 104 };
 105
 106 struct nbd_device {
 107         struct blk_mq_tag_set tag_set;
 108
 109         int index;
 110         refcount_t config_refs;
 111         refcount_t refs;
 112         struct nbd_config *config;
 113         struct mutex config_lock;
 114         struct gendisk *disk;
 115         struct workqueue_struct *recv_workq;
 116
 117         struct list_head list;
 118         struct task_struct *task_recv;
 119         struct task_struct *task_setup;
 120
 121         struct completion *destroy_complete;
 122         unsigned long flags;
 123 };
 124
 125 #define NBD_CMD_REQUEUED        1
 126
 127 struct nbd_cmd {
 128         struct nbd_device *nbd;
 129         struct mutex lock;
 130         int index;
 131         int cookie;
 132         int retries;
 133         blk_status_t status;
 134         unsigned long flags;
 135         u32 cmd_cookie;
 136 };
 137
 138 #if IS_ENABLED(CONFIG_DEBUG_FS)
 139 static struct dentry *nbd_dbg_dir;
 140 #endif
 141
 142 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 143
 144 #define NBD_MAGIC 0x68797548
 145
 146 #define NBD_DEF_BLKSIZE 1024
 147
 148 static unsigned int nbds_max = 16;
 149 static int max_part = 16;
 150 static int part_shift;
 151
 152 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 153 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 154 static void nbd_config_put(struct nbd_device *nbd);
 155 static void nbd_connect_reply(struct genl_info *info, int index);
 156 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 157 static void nbd_dead_link_work(struct work_struct *work);
 158 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 159
 160 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 161 {
 162         return disk_to_dev(nbd->disk);
 163 }
 164
 165 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 166 {
 167         struct request *req = blk_mq_rq_from_pdu(cmd);
 168
 169         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 170                 blk_mq_requeue_request(req, true);
 171 }
 172
 173 #define NBD_COOKIE_BITS 32
 174
 175 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 176 {
 177         struct request *req = blk_mq_rq_from_pdu(cmd);
 178         u32 tag = blk_mq_unique_tag(req);
 179         u64 cookie = cmd->cmd_cookie;
 180
 181         return (cookie << NBD_COOKIE_BITS) | tag;
 182 }
 183
 184 static u32 nbd_handle_to_tag(u64 handle)
 185 {
 186         return (u32)handle;
 187 }
 188
 189 static u32 nbd_handle_to_cookie(u64 handle)
 190 {
 191         return (u32)(handle >> NBD_COOKIE_BITS);
 192 }
 193
 194 static const char *nbdcmd_to_ascii(int cmd)
 195 {
 196         switch (cmd) {
 197         case  NBD_CMD_READ: return "read";
 198         case NBD_CMD_WRITE: return "write";
 199         case  NBD_CMD_DISC: return "disconnect";
 200         case NBD_CMD_FLUSH: return "flush";
 201         case  NBD_CMD_TRIM: return "trim/discard";
 202         }
 203         return "invalid";
 204 }
 205
 206 static ssize_t pid_show(struct device *dev,
 207                         struct device_attribute *attr, char *buf)
 208 {
 209         struct gendisk *disk = dev_to_disk(dev);
 210         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 211
 212         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 213 }
 214
 215 static const struct device_attribute pid_attr = {
 216         .attr = { .name = "pid", .mode = 0444},
 217         .show = pid_show,
 218 };
 219
 220 static void nbd_dev_remove(struct nbd_device *nbd)
 221 {
 222         struct gendisk *disk = nbd->disk;
 223         struct request_queue *q;
 224
 225         if (disk) {
 226                 q = disk->queue;
 227                 del_gendisk(disk);
 228                 blk_cleanup_queue(q);
 229                 blk_mq_free_tag_set(&nbd->tag_set);
 230                 disk->private_data = NULL;
 231                 put_disk(disk);
 232         }
 233
 234         /*
 235          * Place this in the last just before the nbd is freed to
 236          * make sure that the disk and the related kobject are also
 237          * totally removed to avoid duplicate creation of the same
 238          * one.
 239          */
 240         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) && nbd->destroy_complete)
 241                 complete(nbd->destroy_complete);
 242
 243         kfree(nbd);
 244 }
 245
 246 static void nbd_put(struct nbd_device *nbd)
 247 {
 248         if (refcount_dec_and_mutex_lock(&nbd->refs,
 249                                         &nbd_index_mutex)) {
 250                 idr_remove(&nbd_index_idr, nbd->index);
 251                 nbd_dev_remove(nbd);
 252                 mutex_unlock(&nbd_index_mutex);
 253         }
 254 }
 255
 256 static int nbd_disconnected(struct nbd_config *config)
 257 {
 258         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 259                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 260 }
 261
 262 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 263                                 int notify)
 264 {
 265         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 266                 struct link_dead_args *args;
 267                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 268                 if (args) {
 269                         INIT_WORK(&args->work, nbd_dead_link_work);
 270                         args->index = nbd->index;
 271                         queue_work(system_wq, &args->work);
 272                 }
 273         }
 274         if (!nsock->dead) {
 275                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 276                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 277                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 278                                                &nbd->config->runtime_flags)) {
 279                                 set_bit(NBD_RT_DISCONNECTED,
 280                                         &nbd->config->runtime_flags);
 281                                 dev_info(nbd_to_dev(nbd),
 282                                         "Disconnected due to user request.\n");
 283                         }
 284                 }
 285         }
 286         nsock->dead = true;
 287         nsock->pending = NULL;
 288         nsock->sent = 0;
 289 }
 290
 291 static void nbd_size_clear(struct nbd_device *nbd)
 292 {
 293         if (nbd->config->bytesize) {
 294                 set_capacity(nbd->disk, 0);
 295                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 296         }
 297 }
 298
 299 static void nbd_size_update(struct nbd_device *nbd)
 300 {
 301         struct nbd_config *config = nbd->config;
 302         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 303
 304         if (config->flags & NBD_FLAG_SEND_TRIM) {
 305                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 306                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 307                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 308         }
 309         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 310         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 311         set_capacity(nbd->disk, config->bytesize >> 9);
 312         if (bdev) {
 313                 if (bdev->bd_disk) {
 314                         bd_set_size(bdev, config->bytesize);
 315                         set_blocksize(bdev, config->blksize);
 316                 } else
 317                         bdev->bd_invalidated = 1;
 318                 bdput(bdev);
 319         }
 320         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 321 }
 322
 323 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 324                          loff_t nr_blocks)
 325 {
 326         struct nbd_config *config = nbd->config;
 327         config->blksize = blocksize;
 328         config->bytesize = blocksize * nr_blocks;
 329         if (nbd->task_recv != NULL)
 330                 nbd_size_update(nbd);
 331 }
 332
 333 static void nbd_complete_rq(struct request *req)
 334 {
 335         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 336
 337         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 338                 cmd->status ? "failed" : "done");
 339
 340         blk_mq_end_request(req, cmd->status);
 341 }
 342
 343 /*
 344  * Forcibly shutdown the socket causing all listeners to error
 345  */
 346 static void sock_shutdown(struct nbd_device *nbd)
 347 {
 348         struct nbd_config *config = nbd->config;
 349         int i;
 350
 351         if (config->num_connections == 0)
 352                 return;
 353         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 354                 return;
 355
 356         for (i = 0; i < config->num_connections; i++) {
 357                 struct nbd_sock *nsock = config->socks[i];
 358                 mutex_lock(&nsock->tx_lock);
 359                 nbd_mark_nsock_dead(nbd, nsock, 0);
 360                 mutex_unlock(&nsock->tx_lock);
 361         }
 362         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 363 }
 364
 365 static u32 req_to_nbd_cmd_type(struct request *req)
 366 {
 367         switch (req_op(req)) {
 368         case REQ_OP_DISCARD:
 369                 return NBD_CMD_TRIM;
 370         case REQ_OP_FLUSH:
 371                 return NBD_CMD_FLUSH;
 372         case REQ_OP_WRITE:
 373                 return NBD_CMD_WRITE;
 374         case REQ_OP_READ:
 375                 return NBD_CMD_READ;
 376         default:
 377                 return U32_MAX;
 378         }
 379 }
 380
 381 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 382                                                  bool reserved)
 383 {
 384         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 385         struct nbd_device *nbd = cmd->nbd;
 386         struct nbd_config *config;
 387
 388         if (!mutex_trylock(&cmd->lock))
 389                 return BLK_EH_RESET_TIMER;
 390
 391         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 392                 cmd->status = BLK_STS_TIMEOUT;
 393                 mutex_unlock(&cmd->lock);
 394                 goto done;
 395         }
 396         config = nbd->config;
 397
 398         if (config->num_connections > 1 ||
 399             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 400                 dev_err_ratelimited(nbd_to_dev(nbd),
 401                                     "Connection timed out, retrying (%d/%d alive)\n",
 402                                     atomic_read(&config->live_connections),
 403                                     config->num_connections);
 404                 /*
 405                  * Hooray we have more connections, requeue this IO, the submit
 406                  * path will put it on a real connection. Or if only one
 407                  * connection is configured, the submit path will wait util
 408                  * a new connection is reconfigured or util dead timeout.
 409                  */
 410                 if (config->socks) {
 411                         if (cmd->index < config->num_connections) {
 412                                 struct nbd_sock *nsock =
 413                                         config->socks[cmd->index];
 414                                 mutex_lock(&nsock->tx_lock);
 415                                 /* We can have multiple outstanding requests, so
 416                                  * we don't want to mark the nsock dead if we've
 417                                  * already reconnected with a new socket, so
 418                                  * only mark it dead if its the same socket we
 419                                  * were sent out on.
 420                                  */
 421                                 if (cmd->cookie == nsock->cookie)
 422                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 423                                 mutex_unlock(&nsock->tx_lock);
 424                         }
 425                         mutex_unlock(&cmd->lock);
 426                         nbd_requeue_cmd(cmd);
 427                         nbd_config_put(nbd);
 428                         return BLK_EH_DONE;
 429                 }
 430         }
 431
 432         if (!nbd->tag_set.timeout) {
 433                 /*
 434                  * Userspace sets timeout=0 to disable socket disconnection,
 435                  * so just warn and reset the timer.
 436                  */
 437                 struct nbd_sock *nsock = config->socks[cmd->index];
 438                 cmd->retries++;
 439                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 440                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 441                         (unsigned long long)blk_rq_pos(req) << 9,
 442                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 443
 444                 mutex_lock(&nsock->tx_lock);
 445                 if (cmd->cookie != nsock->cookie) {
 446                         nbd_requeue_cmd(cmd);
 447                         mutex_unlock(&nsock->tx_lock);
 448                         mutex_unlock(&cmd->lock);
 449                         nbd_config_put(nbd);
 450                         return BLK_EH_DONE;
 451                 }
 452                 mutex_unlock(&nsock->tx_lock);
 453                 mutex_unlock(&cmd->lock);
 454                 nbd_config_put(nbd);
 455                 return BLK_EH_RESET_TIMER;
 456         }
 457
 458         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 459         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 460         cmd->status = BLK_STS_IOERR;
 461         mutex_unlock(&cmd->lock);
 462         sock_shutdown(nbd);
 463         nbd_config_put(nbd);
 464 done:
 465         blk_mq_complete_request(req);
 466         return BLK_EH_DONE;
 467 }
 468
 469 /*
 470  *  Send or receive packet.
 471  */
 472 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 473                      struct iov_iter *iter, int msg_flags, int *sent)
 474 {
 475         struct nbd_config *config = nbd->config;
 476         struct socket *sock = config->socks[index]->sock;
 477         int result;
 478         struct msghdr msg;
 479         unsigned int noreclaim_flag;
 480
 481         if (unlikely(!sock)) {
 482                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 483                         "Attempted %s on closed socket in sock_xmit\n",
 484                         (send ? "send" : "recv"));
 485                 return -EINVAL;
 486         }
 487
 488         msg.msg_iter = *iter;
 489
 490         noreclaim_flag = memalloc_noreclaim_save();
 491         do {
 492                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 493                 msg.msg_name = NULL;
 494                 msg.msg_namelen = 0;
 495                 msg.msg_control = NULL;
 496                 msg.msg_controllen = 0;
 497                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 498
 499                 if (send)
 500                         result = sock_sendmsg(sock, &msg);
 501                 else
 502                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 503
 504                 if (result <= 0) {
 505                         if (result == 0)
 506                                 result = -EPIPE; /* short read */
 507                         break;
 508                 }
 509                 if (sent)
 510                         *sent += result;
 511         } while (msg_data_left(&msg));
 512
 513         memalloc_noreclaim_restore(noreclaim_flag);
 514
 515         return result;
 516 }
 517
 518 /*
 519  * Different settings for sk->sk_sndtimeo can result in different return values
 520  * if there is a signal pending when we enter sendmsg, because reasons?
 521  */
 522 static inline int was_interrupted(int result)
 523 {
 524         return result == -ERESTARTSYS || result == -EINTR;
 525 }
 526
 527 /* always call with the tx_lock held */
 528 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 529 {
 530         struct request *req = blk_mq_rq_from_pdu(cmd);
 531         struct nbd_config *config = nbd->config;
 532         struct nbd_sock *nsock = config->socks[index];
 533         int result;
 534         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 535         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 536         struct iov_iter from;
 537         unsigned long size = blk_rq_bytes(req);
 538         struct bio *bio;
 539         u64 handle;
 540         u32 type;
 541         u32 nbd_cmd_flags = 0;
 542         int sent = nsock->sent, skip = 0;
 543
 544         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 545
 546         type = req_to_nbd_cmd_type(req);
 547         if (type == U32_MAX)
 548                 return -EIO;
 549
 550         if (rq_data_dir(req) == WRITE &&
 551             (config->flags & NBD_FLAG_READ_ONLY)) {
 552                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 553                                     "Write on read-only\n");
 554                 return -EIO;
 555         }
 556
 557         if (req->cmd_flags & REQ_FUA)
 558                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 559
 560         /* We did a partial send previously, and we at least sent the whole
 561          * request struct, so just go and send the rest of the pages in the
 562          * request.
 563          */
 564         if (sent) {
 565                 if (sent >= sizeof(request)) {
 566                         skip = sent - sizeof(request);
 567
 568                         /* initialize handle for tracing purposes */
 569                         handle = nbd_cmd_handle(cmd);
 570
 571                         goto send_pages;
 572                 }
 573                 iov_iter_advance(&from, sent);
 574         } else {
 575                 cmd->cmd_cookie++;
 576         }
 577         cmd->index = index;
 578         cmd->cookie = nsock->cookie;
 579         cmd->retries = 0;
 580         request.type = htonl(type | nbd_cmd_flags);
 581         if (type != NBD_CMD_FLUSH) {
 582                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 583                 request.len = htonl(size);
 584         }
 585         handle = nbd_cmd_handle(cmd);
 586         memcpy(request.handle, &handle, sizeof(handle));
 587
 588         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 589
 590         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 591                 req, nbdcmd_to_ascii(type),
 592                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 593         result = sock_xmit(nbd, index, 1, &from,
 594                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 595         trace_nbd_header_sent(req, handle);
 596         if (result <= 0) {
 597                 if (was_interrupted(result)) {
 598                         /* If we havne't sent anything we can just return BUSY,
 599                          * however if we have sent something we need to make
 600                          * sure we only allow this req to be sent until we are
 601                          * completely done.
 602                          */
 603                         if (sent) {
 604                                 nsock->pending = req;
 605                                 nsock->sent = sent;
 606                         }
 607                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 608                         return BLK_STS_RESOURCE;
 609                 }
 610                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 611                         "Send control failed (result %d)\n", result);
 612                 return -EAGAIN;
 613         }
 614 send_pages:
 615         if (type != NBD_CMD_WRITE)
 616                 goto out;
 617
 618         bio = req->bio;
 619         while (bio) {
 620                 struct bio *next = bio->bi_next;
 621                 struct bvec_iter iter;
 622                 struct bio_vec bvec;
 623
 624                 bio_for_each_segment(bvec, bio, iter) {
 625                         bool is_last = !next && bio_iter_last(bvec, iter);
 626                         int flags = is_last ? 0 : MSG_MORE;
 627
 628                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 629                                 req, bvec.bv_len);
 630                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 631                         if (skip) {
 632                                 if (skip >= iov_iter_count(&from)) {
 633                                         skip -= iov_iter_count(&from);
 634                                         continue;
 635                                 }
 636                                 iov_iter_advance(&from, skip);
 637                                 skip = 0;
 638                         }
 639                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 640                         if (result <= 0) {
 641                                 if (was_interrupted(result)) {
 642                                         /* We've already sent the header, we
 643                                          * have no choice but to set pending and
 644                                          * return BUSY.
 645                                          */
 646                                         nsock->pending = req;
 647                                         nsock->sent = sent;
 648                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 649                                         return BLK_STS_RESOURCE;
 650                                 }
 651                                 dev_err(disk_to_dev(nbd->disk),
 652                                         "Send data failed (result %d)\n",
 653                                         result);
 654                                 return -EAGAIN;
 655                         }
 656                         /*
 657                          * The completion might already have come in,
 658                          * so break for the last one instead of letting
 659                          * the iterator do it. This prevents use-after-free
 660                          * of the bio.
 661                          */
 662                         if (is_last)
 663                                 break;
 664                 }
 665                 bio = next;
 666         }
 667 out:
 668         trace_nbd_payload_sent(req, handle);
 669         nsock->pending = NULL;
 670         nsock->sent = 0;
 671         return 0;
 672 }
 673
 674 /* NULL returned = something went wrong, inform userspace */
 675 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 676 {
 677         struct nbd_config *config = nbd->config;
 678         int result;
 679         struct nbd_reply reply;
 680         struct nbd_cmd *cmd;
 681         struct request *req = NULL;
 682         u64 handle;
 683         u16 hwq;
 684         u32 tag;
 685         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 686         struct iov_iter to;
 687         int ret = 0;
 688
 689         reply.magic = 0;
 690         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 691         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 692         if (result <= 0) {
 693                 if (!nbd_disconnected(config))
 694                         dev_err(disk_to_dev(nbd->disk),
 695                                 "Receive control failed (result %d)\n", result);
 696                 return ERR_PTR(result);
 697         }
 698
 699         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 700                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 701                                 (unsigned long)ntohl(reply.magic));
 702                 return ERR_PTR(-EPROTO);
 703         }
 704
 705         memcpy(&handle, reply.handle, sizeof(handle));
 706         tag = nbd_handle_to_tag(handle);
 707         hwq = blk_mq_unique_tag_to_hwq(tag);
 708         if (hwq < nbd->tag_set.nr_hw_queues)
 709                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 710                                        blk_mq_unique_tag_to_tag(tag));
 711         if (!req || !blk_mq_request_started(req)) {
 712                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 713                         tag, req);
 714                 return ERR_PTR(-ENOENT);
 715         }
 716         trace_nbd_header_received(req, handle);
 717         cmd = blk_mq_rq_to_pdu(req);
 718
 719         mutex_lock(&cmd->lock);
 720         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 721                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 722                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 723                 ret = -ENOENT;
 724                 goto out;
 725         }
 726         if (cmd->status != BLK_STS_OK) {
 727                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 728                         req);
 729                 ret = -ENOENT;
 730                 goto out;
 731         }
 732         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 733                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 734                         req);
 735                 ret = -ENOENT;
 736                 goto out;
 737         }
 738         if (ntohl(reply.error)) {
 739                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 740                         ntohl(reply.error));
 741                 cmd->status = BLK_STS_IOERR;
 742                 goto out;
 743         }
 744
 745         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 746         if (rq_data_dir(req) != WRITE) {
 747                 struct req_iterator iter;
 748                 struct bio_vec bvec;
 749
 750                 rq_for_each_segment(bvec, req, iter) {
 751                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 752                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 753                         if (result <= 0) {
 754                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 755                                         result);
 756                                 /*
 757                                  * If we've disconnected, we need to make sure we
 758                                  * complete this request, otherwise error out
 759                                  * and let the timeout stuff handle resubmitting
 760                                  * this request onto another connection.
 761                                  */
 762                                 if (nbd_disconnected(config)) {
 763                                         cmd->status = BLK_STS_IOERR;
 764                                         goto out;
 765                                 }
 766                                 ret = -EIO;
 767                                 goto out;
 768                         }
 769                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 770                                 req, bvec.bv_len);
 771                 }
 772         }
 773 out:
 774         trace_nbd_payload_received(req, handle);
 775         mutex_unlock(&cmd->lock);
 776         return ret ? ERR_PTR(ret) : cmd;
 777 }
 778
 779 static void recv_work(struct work_struct *work)
 780 {
 781         struct recv_thread_args *args = container_of(work,
 782                                                      struct recv_thread_args,
 783                                                      work);
 784         struct nbd_device *nbd = args->nbd;
 785         struct nbd_config *config = nbd->config;
 786         struct nbd_cmd *cmd;
 787
 788         while (1) {
 789                 cmd = nbd_read_stat(nbd, args->index);
 790                 if (IS_ERR(cmd)) {
 791                         struct nbd_sock *nsock = config->socks[args->index];
 792
 793                         mutex_lock(&nsock->tx_lock);
 794                         nbd_mark_nsock_dead(nbd, nsock, 1);
 795                         mutex_unlock(&nsock->tx_lock);
 796                         break;
 797                 }
 798
 799                 blk_mq_complete_request(blk_mq_rq_from_pdu(cmd));
 800         }
 801         atomic_dec(&config->recv_threads);
 802         wake_up(&config->recv_wq);
 803         nbd_config_put(nbd);
 804         kfree(args);
 805 }
 806
 807 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 808 {
 809         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 810
 811         mutex_lock(&cmd->lock);
 812         cmd->status = BLK_STS_IOERR;
 813         mutex_unlock(&cmd->lock);
 814
 815         blk_mq_complete_request(req);
 816         return true;
 817 }
 818
 819 static void nbd_clear_que(struct nbd_device *nbd)
 820 {
 821         blk_mq_quiesce_queue(nbd->disk->queue);
 822         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 823         blk_mq_unquiesce_queue(nbd->disk->queue);
 824         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 825 }
 826
 827 static int find_fallback(struct nbd_device *nbd, int index)
 828 {
 829         struct nbd_config *config = nbd->config;
 830         int new_index = -1;
 831         struct nbd_sock *nsock = config->socks[index];
 832         int fallback = nsock->fallback_index;
 833
 834         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 835                 return new_index;
 836
 837         if (config->num_connections <= 1) {
 838                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 839                                     "Dead connection, failed to find a fallback\n");
 840                 return new_index;
 841         }
 842
 843         if (fallback >= 0 && fallback < config->num_connections &&
 844             !config->socks[fallback]->dead)
 845                 return fallback;
 846
 847         if (nsock->fallback_index < 0 ||
 848             nsock->fallback_index >= config->num_connections ||
 849             config->socks[nsock->fallback_index]->dead) {
 850                 int i;
 851                 for (i = 0; i < config->num_connections; i++) {
 852                         if (i == index)
 853                                 continue;
 854                         if (!config->socks[i]->dead) {
 855                                 new_index = i;
 856                                 break;
 857                         }
 858                 }
 859                 nsock->fallback_index = new_index;
 860                 if (new_index < 0) {
 861                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 862                                             "Dead connection, failed to find a fallback\n");
 863                         return new_index;
 864                 }
 865         }
 866         new_index = nsock->fallback_index;
 867         return new_index;
 868 }
 869
 870 static int wait_for_reconnect(struct nbd_device *nbd)
 871 {
 872         struct nbd_config *config = nbd->config;
 873         if (!config->dead_conn_timeout)
 874                 return 0;
 875         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 876                 return 0;
 877         return wait_event_timeout(config->conn_wait,
 878                                   atomic_read(&config->live_connections) > 0,
 879                                   config->dead_conn_timeout) > 0;
 880 }
 881
 882 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 883 {
 884         struct request *req = blk_mq_rq_from_pdu(cmd);
 885         struct nbd_device *nbd = cmd->nbd;
 886         struct nbd_config *config;
 887         struct nbd_sock *nsock;
 888         int ret;
 889
 890         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 891                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 892                                     "Socks array is empty\n");
 893                 blk_mq_start_request(req);
 894                 return -EINVAL;
 895         }
 896         config = nbd->config;
 897
 898         if (index >= config->num_connections) {
 899                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 900                                     "Attempted send on invalid socket\n");
 901                 nbd_config_put(nbd);
 902                 blk_mq_start_request(req);
 903                 return -EINVAL;
 904         }
 905         cmd->status = BLK_STS_OK;
 906 again:
 907         nsock = config->socks[index];
 908         mutex_lock(&nsock->tx_lock);
 909         if (nsock->dead) {
 910                 int old_index = index;
 911                 index = find_fallback(nbd, index);
 912                 mutex_unlock(&nsock->tx_lock);
 913                 if (index < 0) {
 914                         if (wait_for_reconnect(nbd)) {
 915                                 index = old_index;
 916                                 goto again;
 917                         }
 918                         /* All the sockets should already be down at this point,
 919                          * we just want to make sure that DISCONNECTED is set so
 920                          * any requests that come in that were queue'ed waiting
 921                          * for the reconnect timer don't trigger the timer again
 922                          * and instead just error out.
 923                          */
 924                         sock_shutdown(nbd);
 925                         nbd_config_put(nbd);
 926                         blk_mq_start_request(req);
 927                         return -EIO;
 928                 }
 929                 goto again;
 930         }
 931
 932         /* Handle the case that we have a pending request that was partially
 933          * transmitted that _has_ to be serviced first.  We need to call requeue
 934          * here so that it gets put _after_ the request that is already on the
 935          * dispatch list.
 936          */
 937         blk_mq_start_request(req);
 938         if (unlikely(nsock->pending && nsock->pending != req)) {
 939                 nbd_requeue_cmd(cmd);
 940                 ret = 0;
 941                 goto out;
 942         }
 943         /*
 944          * Some failures are related to the link going down, so anything that
 945          * returns EAGAIN can be retried on a different socket.
 946          */
 947         ret = nbd_send_cmd(nbd, cmd, index);
 948         if (ret == -EAGAIN) {
 949                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 950                                     "Request send failed, requeueing\n");
 951                 nbd_mark_nsock_dead(nbd, nsock, 1);
 952                 nbd_requeue_cmd(cmd);
 953                 ret = 0;
 954         }
 955 out:
 956         mutex_unlock(&nsock->tx_lock);
 957         nbd_config_put(nbd);
 958         return ret;
 959 }
 960
 961 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 962                         const struct blk_mq_queue_data *bd)
 963 {
 964         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 965         int ret;
 966
 967         /*
 968          * Since we look at the bio's to send the request over the network we
 969          * need to make sure the completion work doesn't mark this request done
 970          * before we are done doing our send.  This keeps us from dereferencing
 971          * freed data if we have particularly fast completions (ie we get the
 972          * completion before we exit sock_xmit on the last bvec) or in the case
 973          * that the server is misbehaving (or there was an error) before we're
 974          * done sending everything over the wire.
 975          */
 976         mutex_lock(&cmd->lock);
 977         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 978
 979         /* We can be called directly from the user space process, which means we
 980          * could possibly have signals pending so our sendmsg will fail.  In
 981          * this case we need to return that we are busy, otherwise error out as
 982          * appropriate.
 983          */
 984         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 985         if (ret < 0)
 986                 ret = BLK_STS_IOERR;
 987         else if (!ret)
 988                 ret = BLK_STS_OK;
 989         mutex_unlock(&cmd->lock);
 990
 991         return ret;
 992 }
 993
 994 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
 995                                      int *err)
 996 {
 997         struct socket *sock;
 998
 999         *err = 0;
1000         sock = sockfd_lookup(fd, err);
1001         if (!sock)
1002                 return NULL;
1003
1004         if (sock->ops->shutdown == sock_no_shutdown) {
1005                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1006                 *err = -EINVAL;
1007                 sockfd_put(sock);
1008                 return NULL;
1009         }
1010
1011         return sock;
1012 }
1013
1014 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1015                           bool netlink)
1016 {
1017         struct nbd_config *config = nbd->config;
1018         struct socket *sock;
1019         struct nbd_sock **socks;
1020         struct nbd_sock *nsock;
1021         int err;
1022
1023         sock = nbd_get_socket(nbd, arg, &err);
1024         if (!sock)
1025                 return err;
1026
1027         if (!netlink && !nbd->task_setup &&
1028             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1029                 nbd->task_setup = current;
1030
1031         if (!netlink &&
1032             (nbd->task_setup != current ||
1033              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1034                 dev_err(disk_to_dev(nbd->disk),
1035                         "Device being setup by another task");
1036                 err = -EBUSY;
1037                 goto put_socket;
1038         }
1039
1040         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1041         if (!nsock) {
1042                 err = -ENOMEM;
1043                 goto put_socket;
1044         }
1045
1046         socks = krealloc(config->socks, (config->num_connections + 1) *
1047                          sizeof(struct nbd_sock *), GFP_KERNEL);
1048         if (!socks) {
1049                 kfree(nsock);
1050                 err = -ENOMEM;
1051                 goto put_socket;
1052         }
1053
1054         config->socks = socks;
1055
1056         nsock->fallback_index = -1;
1057         nsock->dead = false;
1058         mutex_init(&nsock->tx_lock);
1059         nsock->sock = sock;
1060         nsock->pending = NULL;
1061         nsock->sent = 0;
1062         nsock->cookie = 0;
1063         socks[config->num_connections++] = nsock;
1064         atomic_inc(&config->live_connections);
1065
1066         return 0;
1067
1068 put_socket:
1069         sockfd_put(sock);
1070         return err;
1071 }
1072
1073 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1074 {
1075         struct nbd_config *config = nbd->config;
1076         struct socket *sock, *old;
1077         struct recv_thread_args *args;
1078         int i;
1079         int err;
1080
1081         sock = nbd_get_socket(nbd, arg, &err);
1082         if (!sock)
1083                 return err;
1084
1085         args = kzalloc(sizeof(*args), GFP_KERNEL);
1086         if (!args) {
1087                 sockfd_put(sock);
1088                 return -ENOMEM;
1089         }
1090
1091         for (i = 0; i < config->num_connections; i++) {
1092                 struct nbd_sock *nsock = config->socks[i];
1093
1094                 if (!nsock->dead)
1095                         continue;
1096
1097                 mutex_lock(&nsock->tx_lock);
1098                 if (!nsock->dead) {
1099                         mutex_unlock(&nsock->tx_lock);
1100                         continue;
1101                 }
1102                 sk_set_memalloc(sock->sk);
1103                 if (nbd->tag_set.timeout)
1104                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1105                 atomic_inc(&config->recv_threads);
1106                 refcount_inc(&nbd->config_refs);
1107                 old = nsock->sock;
1108                 nsock->fallback_index = -1;
1109                 nsock->sock = sock;
1110                 nsock->dead = false;
1111                 INIT_WORK(&args->work, recv_work);
1112                 args->index = i;
1113                 args->nbd = nbd;
1114                 nsock->cookie++;
1115                 mutex_unlock(&nsock->tx_lock);
1116                 sockfd_put(old);
1117
1118                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1119
1120                 /* We take the tx_mutex in an error path in the recv_work, so we
1121                  * need to queue_work outside of the tx_mutex.
1122                  */
1123                 queue_work(nbd->recv_workq, &args->work);
1124
1125                 atomic_inc(&config->live_connections);
1126                 wake_up(&config->conn_wait);
1127                 return 0;
1128         }
1129         sockfd_put(sock);
1130         kfree(args);
1131         return -ENOSPC;
1132 }
1133
1134 static void nbd_bdev_reset(struct block_device *bdev)
1135 {
1136         if (bdev->bd_openers > 1)
1137                 return;
1138         bd_set_size(bdev, 0);
1139 }
1140
1141 static void nbd_parse_flags(struct nbd_device *nbd)
1142 {
1143         struct nbd_config *config = nbd->config;
1144         if (config->flags & NBD_FLAG_READ_ONLY)
1145                 set_disk_ro(nbd->disk, true);
1146         else
1147                 set_disk_ro(nbd->disk, false);
1148         if (config->flags & NBD_FLAG_SEND_TRIM)
1149                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1150         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1151                 if (config->flags & NBD_FLAG_SEND_FUA)
1152                         blk_queue_write_cache(nbd->disk->queue, true, true);
1153                 else
1154                         blk_queue_write_cache(nbd->disk->queue, true, false);
1155         }
1156         else
1157                 blk_queue_write_cache(nbd->disk->queue, false, false);
1158 }
1159
1160 static void send_disconnects(struct nbd_device *nbd)
1161 {
1162         struct nbd_config *config = nbd->config;
1163         struct nbd_request request = {
1164                 .magic = htonl(NBD_REQUEST_MAGIC),
1165                 .type = htonl(NBD_CMD_DISC),
1166         };
1167         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1168         struct iov_iter from;
1169         int i, ret;
1170
1171         for (i = 0; i < config->num_connections; i++) {
1172                 struct nbd_sock *nsock = config->socks[i];
1173
1174                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1175                 mutex_lock(&nsock->tx_lock);
1176                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1177                 if (ret <= 0)
1178                         dev_err(disk_to_dev(nbd->disk),
1179                                 "Send disconnect failed %d\n", ret);
1180                 mutex_unlock(&nsock->tx_lock);
1181         }
1182 }
1183
1184 static int nbd_disconnect(struct nbd_device *nbd)
1185 {
1186         struct nbd_config *config = nbd->config;
1187
1188         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1189         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1190         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1191         send_disconnects(nbd);
1192         return 0;
1193 }
1194
1195 static void nbd_clear_sock(struct nbd_device *nbd)
1196 {
1197         sock_shutdown(nbd);
1198         nbd_clear_que(nbd);
1199         nbd->task_setup = NULL;
1200 }
1201
1202 static void nbd_config_put(struct nbd_device *nbd)
1203 {
1204         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1205                                         &nbd->config_lock)) {
1206                 struct nbd_config *config = nbd->config;
1207                 nbd_dev_dbg_close(nbd);
1208                 nbd_size_clear(nbd);
1209                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1210                                        &config->runtime_flags))
1211                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1212                 nbd->task_recv = NULL;
1213                 nbd_clear_sock(nbd);
1214                 if (config->num_connections) {
1215                         int i;
1216                         for (i = 0; i < config->num_connections; i++) {
1217                                 sockfd_put(config->socks[i]->sock);
1218                                 kfree(config->socks[i]);
1219                         }
1220                         kfree(config->socks);
1221                 }
1222                 kfree(nbd->config);
1223                 nbd->config = NULL;
1224
1225                 if (nbd->recv_workq)
1226                         destroy_workqueue(nbd->recv_workq);
1227                 nbd->recv_workq = NULL;
1228
1229                 nbd->tag_set.timeout = 0;
1230                 nbd->disk->queue->limits.discard_granularity = 0;
1231                 nbd->disk->queue->limits.discard_alignment = 0;
1232                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1233                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1234
1235                 mutex_unlock(&nbd->config_lock);
1236                 nbd_put(nbd);
1237                 module_put(THIS_MODULE);
1238         }
1239 }
1240
1241 static int nbd_start_device(struct nbd_device *nbd)
1242 {
1243         struct nbd_config *config = nbd->config;
1244         int num_connections = config->num_connections;
1245         int error = 0, i;
1246
1247         if (nbd->task_recv)
1248                 return -EBUSY;
1249         if (!config->socks)
1250                 return -EINVAL;
1251         if (num_connections > 1 &&
1252             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1253                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1254                 return -EINVAL;
1255         }
1256
1257         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1258                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1259                                           WQ_UNBOUND, 0, nbd->index);
1260         if (!nbd->recv_workq) {
1261                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1262                 return -ENOMEM;
1263         }
1264
1265         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1266         nbd->task_recv = current;
1267
1268         nbd_parse_flags(nbd);
1269
1270         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1271         if (error) {
1272                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1273                 return error;
1274         }
1275         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1276
1277         nbd_dev_dbg_init(nbd);
1278         for (i = 0; i < num_connections; i++) {
1279                 struct recv_thread_args *args;
1280
1281                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1282                 if (!args) {
1283                         sock_shutdown(nbd);
1284                         /*
1285                          * If num_connections is m (2 < m),
1286                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1287                          * But NO.(n + 1) failed. We still have n recv threads.
1288                          * So, add flush_workqueue here to prevent recv threads
1289                          * dropping the last config_refs and trying to destroy
1290                          * the workqueue from inside the workqueue.
1291                          */
1292                         if (i)
1293                                 flush_workqueue(nbd->recv_workq);
1294                         return -ENOMEM;
1295                 }
1296                 sk_set_memalloc(config->socks[i]->sock->sk);
1297                 if (nbd->tag_set.timeout)
1298                         config->socks[i]->sock->sk->sk_sndtimeo =
1299                                 nbd->tag_set.timeout;
1300                 atomic_inc(&config->recv_threads);
1301                 refcount_inc(&nbd->config_refs);
1302                 INIT_WORK(&args->work, recv_work);
1303                 args->nbd = nbd;
1304                 args->index = i;
1305                 queue_work(nbd->recv_workq, &args->work);
1306         }
1307         nbd_size_update(nbd);
1308         return error;
1309 }
1310
1311 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1312 {
1313         struct nbd_config *config = nbd->config;
1314         int ret;
1315
1316         ret = nbd_start_device(nbd);
1317         if (ret)
1318                 return ret;
1319
1320         if (max_part)
1321                 bdev->bd_invalidated = 1;
1322         mutex_unlock(&nbd->config_lock);
1323         ret = wait_event_interruptible(config->recv_wq,
1324                                          atomic_read(&config->recv_threads) == 0);
1325         if (ret)
1326                 sock_shutdown(nbd);
1327         flush_workqueue(nbd->recv_workq);
1328
1329         mutex_lock(&nbd->config_lock);
1330         nbd_bdev_reset(bdev);
1331         /* user requested, ignore socket errors */
1332         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1333                 ret = 0;
1334         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1335                 ret = -ETIMEDOUT;
1336         return ret;
1337 }
1338
1339 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1340                                  struct block_device *bdev)
1341 {
1342         sock_shutdown(nbd);
1343         __invalidate_device(bdev, true);
1344         nbd_bdev_reset(bdev);
1345         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1346                                &nbd->config->runtime_flags))
1347                 nbd_config_put(nbd);
1348 }
1349
1350 static bool nbd_is_valid_blksize(unsigned long blksize)
1351 {
1352         if (!blksize || !is_power_of_2(blksize) || blksize < 512 ||
1353             blksize > PAGE_SIZE)
1354                 return false;
1355         return true;
1356 }
1357
1358 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1359 {
1360         nbd->tag_set.timeout = timeout * HZ;
1361         if (timeout)
1362                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1363 }
1364
1365 /* Must be called with config_lock held */
1366 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1367                        unsigned int cmd, unsigned long arg)
1368 {
1369         struct nbd_config *config = nbd->config;
1370
1371         switch (cmd) {
1372         case NBD_DISCONNECT:
1373                 return nbd_disconnect(nbd);
1374         case NBD_CLEAR_SOCK:
1375                 nbd_clear_sock_ioctl(nbd, bdev);
1376                 return 0;
1377         case NBD_SET_SOCK:
1378                 return nbd_add_socket(nbd, arg, false);
1379         case NBD_SET_BLKSIZE:
1380                 if (!arg)
1381                         arg = NBD_DEF_BLKSIZE;
1382                 if (!nbd_is_valid_blksize(arg))
1383                         return -EINVAL;
1384                 nbd_size_set(nbd, arg,
1385                              div_s64(config->bytesize, arg));
1386                 return 0;
1387         case NBD_SET_SIZE:
1388                 nbd_size_set(nbd, config->blksize,
1389                              div_s64(arg, config->blksize));
1390                 return 0;
1391         case NBD_SET_SIZE_BLOCKS:
1392                 nbd_size_set(nbd, config->blksize, arg);
1393                 return 0;
1394         case NBD_SET_TIMEOUT:
1395                 nbd_set_cmd_timeout(nbd, arg);
1396                 return 0;
1397
1398         case NBD_SET_FLAGS:
1399                 config->flags = arg;
1400                 return 0;
1401         case NBD_DO_IT:
1402                 return nbd_start_device_ioctl(nbd, bdev);
1403         case NBD_CLEAR_QUE:
1404                 /*
1405                  * This is for compatibility only.  The queue is always cleared
1406                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1407                  */
1408                 return 0;
1409         case NBD_PRINT_DEBUG:
1410                 /*
1411                  * For compatibility only, we no longer keep a list of
1412                  * outstanding requests.
1413                  */
1414                 return 0;
1415         }
1416         return -ENOTTY;
1417 }
1418
1419 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1420                      unsigned int cmd, unsigned long arg)
1421 {
1422         struct nbd_device *nbd = bdev->bd_disk->private_data;
1423         struct nbd_config *config = nbd->config;
1424         int error = -EINVAL;
1425
1426         if (!capable(CAP_SYS_ADMIN))
1427                 return -EPERM;
1428
1429         /* The block layer will pass back some non-nbd ioctls in case we have
1430          * special handling for them, but we don't so just return an error.
1431          */
1432         if (_IOC_TYPE(cmd) != 0xab)
1433                 return -EINVAL;
1434
1435         mutex_lock(&nbd->config_lock);
1436
1437         /* Don't allow ioctl operations on a nbd device that was created with
1438          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1439          */
1440         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1441             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1442                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1443         else
1444                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1445         mutex_unlock(&nbd->config_lock);
1446         return error;
1447 }
1448
1449 static struct nbd_config *nbd_alloc_config(void)
1450 {
1451         struct nbd_config *config;
1452
1453         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1454         if (!config)
1455                 return NULL;
1456         atomic_set(&config->recv_threads, 0);
1457         init_waitqueue_head(&config->recv_wq);
1458         init_waitqueue_head(&config->conn_wait);
1459         config->blksize = NBD_DEF_BLKSIZE;
1460         atomic_set(&config->live_connections, 0);
1461         try_module_get(THIS_MODULE);
1462         return config;
1463 }
1464
1465 static int nbd_open(struct block_device *bdev, fmode_t mode)
1466 {
1467         struct nbd_device *nbd;
1468         int ret = 0;
1469
1470         mutex_lock(&nbd_index_mutex);
1471         nbd = bdev->bd_disk->private_data;
1472         if (!nbd) {
1473                 ret = -ENXIO;
1474                 goto out;
1475         }
1476         if (!refcount_inc_not_zero(&nbd->refs)) {
1477                 ret = -ENXIO;
1478                 goto out;
1479         }
1480         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1481                 struct nbd_config *config;
1482
1483                 mutex_lock(&nbd->config_lock);
1484                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1485                         mutex_unlock(&nbd->config_lock);
1486                         goto out;
1487                 }
1488                 config = nbd->config = nbd_alloc_config();
1489                 if (!config) {
1490                         ret = -ENOMEM;
1491                         mutex_unlock(&nbd->config_lock);
1492                         goto out;
1493                 }
1494                 refcount_set(&nbd->config_refs, 1);
1495                 refcount_inc(&nbd->refs);
1496                 mutex_unlock(&nbd->config_lock);
1497                 bdev->bd_invalidated = 1;
1498         } else if (nbd_disconnected(nbd->config)) {
1499                 bdev->bd_invalidated = 1;
1500         }
1501 out:
1502         mutex_unlock(&nbd_index_mutex);
1503         return ret;
1504 }
1505
1506 static void nbd_release(struct gendisk *disk, fmode_t mode)
1507 {
1508         struct nbd_device *nbd = disk->private_data;
1509         struct block_device *bdev = bdget_disk(disk, 0);
1510
1511         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1512                         bdev->bd_openers == 0)
1513                 nbd_disconnect_and_put(nbd);
1514
1515         nbd_config_put(nbd);
1516         nbd_put(nbd);
1517 }
1518
1519 static const struct block_device_operations nbd_fops =
1520 {
1521         .owner =        THIS_MODULE,
1522         .open =         nbd_open,
1523         .release =      nbd_release,
1524         .ioctl =        nbd_ioctl,
1525         .compat_ioctl = nbd_ioctl,
1526 };
1527
1528 #if IS_ENABLED(CONFIG_DEBUG_FS)
1529
1530 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1531 {
1532         struct nbd_device *nbd = s->private;
1533
1534         if (nbd->task_recv)
1535                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1536
1537         return 0;
1538 }
1539
1540 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1541 {
1542         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1543 }
1544
1545 static const struct file_operations nbd_dbg_tasks_ops = {
1546         .open = nbd_dbg_tasks_open,
1547         .read = seq_read,
1548         .llseek = seq_lseek,
1549         .release = single_release,
1550 };
1551
1552 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1553 {
1554         struct nbd_device *nbd = s->private;
1555         u32 flags = nbd->config->flags;
1556
1557         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1558
1559         seq_puts(s, "Known flags:\n");
1560
1561         if (flags & NBD_FLAG_HAS_FLAGS)
1562                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1563         if (flags & NBD_FLAG_READ_ONLY)
1564                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1565         if (flags & NBD_FLAG_SEND_FLUSH)
1566                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1567         if (flags & NBD_FLAG_SEND_FUA)
1568                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1569         if (flags & NBD_FLAG_SEND_TRIM)
1570                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1571
1572         return 0;
1573 }
1574
1575 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1576 {
1577         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1578 }
1579
1580 static const struct file_operations nbd_dbg_flags_ops = {
1581         .open = nbd_dbg_flags_open,
1582         .read = seq_read,
1583         .llseek = seq_lseek,
1584         .release = single_release,
1585 };
1586
1587 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1588 {
1589         struct dentry *dir;
1590         struct nbd_config *config = nbd->config;
1591
1592         if (!nbd_dbg_dir)
1593                 return -EIO;
1594
1595         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1596         if (!dir) {
1597                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1598                         nbd_name(nbd));
1599                 return -EIO;
1600         }
1601         config->dbg_dir = dir;
1602
1603         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1604         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1605         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1606         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1607         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1608
1609         return 0;
1610 }
1611
1612 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1613 {
1614         debugfs_remove_recursive(nbd->config->dbg_dir);
1615 }
1616
1617 static int nbd_dbg_init(void)
1618 {
1619         struct dentry *dbg_dir;
1620
1621         dbg_dir = debugfs_create_dir("nbd", NULL);
1622         if (!dbg_dir)
1623                 return -EIO;
1624
1625         nbd_dbg_dir = dbg_dir;
1626
1627         return 0;
1628 }
1629
1630 static void nbd_dbg_close(void)
1631 {
1632         debugfs_remove_recursive(nbd_dbg_dir);
1633 }
1634
1635 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1636
1637 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1638 {
1639         return 0;
1640 }
1641
1642 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1643 {
1644 }
1645
1646 static int nbd_dbg_init(void)
1647 {
1648         return 0;
1649 }
1650
1651 static void nbd_dbg_close(void)
1652 {
1653 }
1654
1655 #endif
1656
1657 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1658                             unsigned int hctx_idx, unsigned int numa_node)
1659 {
1660         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1661         cmd->nbd = set->driver_data;
1662         cmd->flags = 0;
1663         mutex_init(&cmd->lock);
1664         return 0;
1665 }
1666
1667 static const struct blk_mq_ops nbd_mq_ops = {
1668         .queue_rq       = nbd_queue_rq,
1669         .complete       = nbd_complete_rq,
1670         .init_request   = nbd_init_request,
1671         .timeout        = nbd_xmit_timeout,
1672 };
1673
1674 static int nbd_dev_add(int index)
1675 {
1676         struct nbd_device *nbd;
1677         struct gendisk *disk;
1678         struct request_queue *q;
1679         int err = -ENOMEM;
1680
1681         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1682         if (!nbd)
1683                 goto out;
1684
1685         disk = alloc_disk(1 << part_shift);
1686         if (!disk)
1687                 goto out_free_nbd;
1688
1689         if (index >= 0) {
1690                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1691                                 GFP_KERNEL);
1692                 if (err == -ENOSPC)
1693                         err = -EEXIST;
1694         } else {
1695                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1696                 if (err >= 0)
1697                         index = err;
1698         }
1699         if (err < 0)
1700                 goto out_free_disk;
1701
1702         nbd->index = index;
1703         nbd->disk = disk;
1704         nbd->tag_set.ops = &nbd_mq_ops;
1705         nbd->tag_set.nr_hw_queues = 1;
1706         nbd->tag_set.queue_depth = 128;
1707         nbd->tag_set.numa_node = NUMA_NO_NODE;
1708         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1709         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1710                 BLK_MQ_F_BLOCKING;
1711         nbd->tag_set.driver_data = nbd;
1712         nbd->destroy_complete = NULL;
1713
1714         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1715         if (err)
1716                 goto out_free_idr;
1717
1718         q = blk_mq_init_queue(&nbd->tag_set);
1719         if (IS_ERR(q)) {
1720                 err = PTR_ERR(q);
1721                 goto out_free_tags;
1722         }
1723         disk->queue = q;
1724
1725         /*
1726          * Tell the block layer that we are not a rotational device
1727          */
1728         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1729         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1730         disk->queue->limits.discard_granularity = 0;
1731         disk->queue->limits.discard_alignment = 0;
1732         blk_queue_max_discard_sectors(disk->queue, 0);
1733         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1734         blk_queue_max_segments(disk->queue, USHRT_MAX);
1735         blk_queue_max_hw_sectors(disk->queue, 65536);
1736         disk->queue->limits.max_sectors = 256;
1737
1738         mutex_init(&nbd->config_lock);
1739         refcount_set(&nbd->config_refs, 0);
1740         refcount_set(&nbd->refs, 1);
1741         INIT_LIST_HEAD(&nbd->list);
1742         disk->major = NBD_MAJOR;
1743         disk->first_minor = index << part_shift;
1744         disk->fops = &nbd_fops;
1745         disk->private_data = nbd;
1746         sprintf(disk->disk_name, "nbd%d", index);
1747         add_disk(disk);
1748         nbd_total_devices++;
1749         return index;
1750
1751 out_free_tags:
1752         blk_mq_free_tag_set(&nbd->tag_set);
1753 out_free_idr:
1754         idr_remove(&nbd_index_idr, index);
1755 out_free_disk:
1756         put_disk(disk);
1757 out_free_nbd:
1758         kfree(nbd);
1759 out:
1760         return err;
1761 }
1762
1763 static int find_free_cb(int id, void *ptr, void *data)
1764 {
1765         struct nbd_device *nbd = ptr;
1766         struct nbd_device **found = data;
1767
1768         if (!refcount_read(&nbd->config_refs)) {
1769                 *found = nbd;
1770                 return 1;
1771         }
1772         return 0;
1773 }
1774
1775 /* Netlink interface. */
1776 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1777         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1778         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1779         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1780         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1781         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1782         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1783         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1784         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1785         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1786 };
1787
1788 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1789         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1790 };
1791
1792 /* We don't use this right now since we don't parse the incoming list, but we
1793  * still want it here so userspace knows what to expect.
1794  */
1795 static const struct nla_policy __attribute__((unused))
1796 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1797         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1798         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1799 };
1800
1801 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1802 {
1803         struct nbd_config *config = nbd->config;
1804         u64 bsize = config->blksize;
1805         u64 bytes = config->bytesize;
1806
1807         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1808                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1809
1810         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1811                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1812                 if (!bsize)
1813                         bsize = NBD_DEF_BLKSIZE;
1814                 if (!nbd_is_valid_blksize(bsize)) {
1815                         printk(KERN_ERR "Invalid block size %llu\n", bsize);
1816                         return -EINVAL;
1817                 }
1818         }
1819
1820         if (bytes != config->bytesize || bsize != config->blksize)
1821                 nbd_size_set(nbd, bsize, div64_u64(bytes, bsize));
1822         return 0;
1823 }
1824
1825 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1826 {
1827         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1828         struct nbd_device *nbd = NULL;
1829         struct nbd_config *config;
1830         int index = -1;
1831         int ret;
1832         bool put_dev = false;
1833
1834         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1835                 return -EPERM;
1836
1837         if (info->attrs[NBD_ATTR_INDEX])
1838                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1839         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1840                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1841                 return -EINVAL;
1842         }
1843         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1844                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1845                 return -EINVAL;
1846         }
1847 again:
1848         mutex_lock(&nbd_index_mutex);
1849         if (index == -1) {
1850                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1851                 if (ret == 0) {
1852                         int new_index;
1853                         new_index = nbd_dev_add(-1);
1854                         if (new_index < 0) {
1855                                 mutex_unlock(&nbd_index_mutex);
1856                                 printk(KERN_ERR "nbd: failed to add new device\n");
1857                                 return new_index;
1858                         }
1859                         nbd = idr_find(&nbd_index_idr, new_index);
1860                 }
1861         } else {
1862                 nbd = idr_find(&nbd_index_idr, index);
1863                 if (!nbd) {
1864                         ret = nbd_dev_add(index);
1865                         if (ret < 0) {
1866                                 mutex_unlock(&nbd_index_mutex);
1867                                 printk(KERN_ERR "nbd: failed to add new device\n");
1868                                 return ret;
1869                         }
1870                         nbd = idr_find(&nbd_index_idr, index);
1871                 }
1872         }
1873         if (!nbd) {
1874                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1875                        index);
1876                 mutex_unlock(&nbd_index_mutex);
1877                 return -EINVAL;
1878         }
1879
1880         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1881             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1882                 nbd->destroy_complete = &destroy_complete;
1883                 mutex_unlock(&nbd_index_mutex);
1884
1885                 /* Wait untill the the nbd stuff is totally destroyed */
1886                 wait_for_completion(&destroy_complete);
1887                 goto again;
1888         }
1889
1890         if (!refcount_inc_not_zero(&nbd->refs)) {
1891                 mutex_unlock(&nbd_index_mutex);
1892                 if (index == -1)
1893                         goto again;
1894                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1895                        index);
1896                 return -EINVAL;
1897         }
1898         mutex_unlock(&nbd_index_mutex);
1899
1900         mutex_lock(&nbd->config_lock);
1901         if (refcount_read(&nbd->config_refs)) {
1902                 mutex_unlock(&nbd->config_lock);
1903                 nbd_put(nbd);
1904                 if (index == -1)
1905                         goto again;
1906                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1907                 return -EBUSY;
1908         }
1909         if (WARN_ON(nbd->config)) {
1910                 mutex_unlock(&nbd->config_lock);
1911                 nbd_put(nbd);
1912                 return -EINVAL;
1913         }
1914         config = nbd->config = nbd_alloc_config();
1915         if (!nbd->config) {
1916                 mutex_unlock(&nbd->config_lock);
1917                 nbd_put(nbd);
1918                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1919                 return -ENOMEM;
1920         }
1921         refcount_set(&nbd->config_refs, 1);
1922         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1923
1924         ret = nbd_genl_size_set(info, nbd);
1925         if (ret)
1926                 goto out;
1927
1928         if (info->attrs[NBD_ATTR_TIMEOUT])
1929                 nbd_set_cmd_timeout(nbd,
1930                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1931         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1932                 config->dead_conn_timeout =
1933                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1934                 config->dead_conn_timeout *= HZ;
1935         }
1936         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1937                 config->flags =
1938                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1939         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1940                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1941                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1942                         set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
1943                                 &config->runtime_flags);
1944                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1945                         put_dev = true;
1946                 } else {
1947                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1948                 }
1949                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1950                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1951                                 &config->runtime_flags);
1952                 }
1953         }
1954
1955         if (info->attrs[NBD_ATTR_SOCKETS]) {
1956                 struct nlattr *attr;
1957                 int rem, fd;
1958
1959                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1960                                     rem) {
1961                         struct nlattr *socks[NBD_SOCK_MAX+1];
1962
1963                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1964                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1965                                 ret = -EINVAL;
1966                                 goto out;
1967                         }
1968                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1969                                                           attr,
1970                                                           nbd_sock_policy,
1971                                                           info->extack);
1972                         if (ret != 0) {
1973                                 printk(KERN_ERR "nbd: error processing sock list\n");
1974                                 ret = -EINVAL;
1975                                 goto out;
1976                         }
1977                         if (!socks[NBD_SOCK_FD])
1978                                 continue;
1979                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1980                         ret = nbd_add_socket(nbd, fd, true);
1981                         if (ret)
1982                                 goto out;
1983                 }
1984         }
1985         ret = nbd_start_device(nbd);
1986 out:
1987         mutex_unlock(&nbd->config_lock);
1988         if (!ret) {
1989                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
1990                 refcount_inc(&nbd->config_refs);
1991                 nbd_connect_reply(info, nbd->index);
1992         }
1993         nbd_config_put(nbd);
1994         if (put_dev)
1995                 nbd_put(nbd);
1996         return ret;
1997 }
1998
1999 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2000 {
2001         mutex_lock(&nbd->config_lock);
2002         nbd_disconnect(nbd);
2003         nbd_clear_sock(nbd);
2004         mutex_unlock(&nbd->config_lock);
2005         /*
2006          * Make sure recv thread has finished, so it does not drop the last
2007          * config ref and try to destroy the workqueue from inside the work
2008          * queue.
2009          */
2010         flush_workqueue(nbd->recv_workq);
2011         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2012                                &nbd->config->runtime_flags))
2013                 nbd_config_put(nbd);
2014 }
2015
2016 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2017 {
2018         struct nbd_device *nbd;
2019         int index;
2020
2021         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2022                 return -EPERM;
2023
2024         if (!info->attrs[NBD_ATTR_INDEX]) {
2025                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2026                 return -EINVAL;
2027         }
2028         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2029         mutex_lock(&nbd_index_mutex);
2030         nbd = idr_find(&nbd_index_idr, index);
2031         if (!nbd) {
2032                 mutex_unlock(&nbd_index_mutex);
2033                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2034                        index);
2035                 return -EINVAL;
2036         }
2037         if (!refcount_inc_not_zero(&nbd->refs)) {
2038                 mutex_unlock(&nbd_index_mutex);
2039                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2040                        index);
2041                 return -EINVAL;
2042         }
2043         mutex_unlock(&nbd_index_mutex);
2044         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2045                 nbd_put(nbd);
2046                 return 0;
2047         }
2048         nbd_disconnect_and_put(nbd);
2049         nbd_config_put(nbd);
2050         nbd_put(nbd);
2051         return 0;
2052 }
2053
2054 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2055 {
2056         struct nbd_device *nbd = NULL;
2057         struct nbd_config *config;
2058         int index;
2059         int ret = 0;
2060         bool put_dev = false;
2061
2062         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2063                 return -EPERM;
2064
2065         if (!info->attrs[NBD_ATTR_INDEX]) {
2066                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2067                 return -EINVAL;
2068         }
2069         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2070         mutex_lock(&nbd_index_mutex);
2071         nbd = idr_find(&nbd_index_idr, index);
2072         if (!nbd) {
2073                 mutex_unlock(&nbd_index_mutex);
2074                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2075                        index);
2076                 return -EINVAL;
2077         }
2078         if (!refcount_inc_not_zero(&nbd->refs)) {
2079                 mutex_unlock(&nbd_index_mutex);
2080                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2081                        index);
2082                 return -EINVAL;
2083         }
2084         mutex_unlock(&nbd_index_mutex);
2085
2086         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2087                 dev_err(nbd_to_dev(nbd),
2088                         "not configured, cannot reconfigure\n");
2089                 nbd_put(nbd);
2090                 return -EINVAL;
2091         }
2092
2093         mutex_lock(&nbd->config_lock);
2094         config = nbd->config;
2095         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2096             !nbd->task_recv) {
2097                 dev_err(nbd_to_dev(nbd),
2098                         "not configured, cannot reconfigure\n");
2099                 ret = -EINVAL;
2100                 goto out;
2101         }
2102
2103         ret = nbd_genl_size_set(info, nbd);
2104         if (ret)
2105                 goto out;
2106
2107         if (info->attrs[NBD_ATTR_TIMEOUT])
2108                 nbd_set_cmd_timeout(nbd,
2109                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2110         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2111                 config->dead_conn_timeout =
2112                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2113                 config->dead_conn_timeout *= HZ;
2114         }
2115         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2116                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2117                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2118                         if (!test_and_set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2119                                               &config->runtime_flags))
2120                                 put_dev = true;
2121                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2122                 } else {
2123                         if (test_and_clear_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2124                                                &config->runtime_flags))
2125                                 refcount_inc(&nbd->refs);
2126                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2127                 }
2128
2129                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2130                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2131                                         &config->runtime_flags);
2132                 } else {
2133                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2134                                         &config->runtime_flags);
2135                 }
2136         }
2137
2138         if (info->attrs[NBD_ATTR_SOCKETS]) {
2139                 struct nlattr *attr;
2140                 int rem, fd;
2141
2142                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2143                                     rem) {
2144                         struct nlattr *socks[NBD_SOCK_MAX+1];
2145
2146                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2147                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2148                                 ret = -EINVAL;
2149                                 goto out;
2150                         }
2151                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2152                                                           attr,
2153                                                           nbd_sock_policy,
2154                                                           info->extack);
2155                         if (ret != 0) {
2156                                 printk(KERN_ERR "nbd: error processing sock list\n");
2157                                 ret = -EINVAL;
2158                                 goto out;
2159                         }
2160                         if (!socks[NBD_SOCK_FD])
2161                                 continue;
2162                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2163                         ret = nbd_reconnect_socket(nbd, fd);
2164                         if (ret) {
2165                                 if (ret == -ENOSPC)
2166                                         ret = 0;
2167                                 goto out;
2168                         }
2169                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2170                 }
2171         }
2172 out:
2173         mutex_unlock(&nbd->config_lock);
2174         nbd_config_put(nbd);
2175         nbd_put(nbd);
2176         if (put_dev)
2177                 nbd_put(nbd);
2178         return ret;
2179 }
2180
2181 static const struct genl_ops nbd_connect_genl_ops[] = {
2182         {
2183                 .cmd    = NBD_CMD_CONNECT,
2184                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2185                 .doit   = nbd_genl_connect,
2186         },
2187         {
2188                 .cmd    = NBD_CMD_DISCONNECT,
2189                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2190                 .doit   = nbd_genl_disconnect,
2191         },
2192         {
2193                 .cmd    = NBD_CMD_RECONFIGURE,
2194                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2195                 .doit   = nbd_genl_reconfigure,
2196         },
2197         {
2198                 .cmd    = NBD_CMD_STATUS,
2199                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2200                 .doit   = nbd_genl_status,
2201         },
2202 };
2203
2204 static const struct genl_multicast_group nbd_mcast_grps[] = {
2205         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2206 };
2207
2208 static struct genl_family nbd_genl_family __ro_after_init = {
2209         .hdrsize        = 0,
2210         .name           = NBD_GENL_FAMILY_NAME,
2211         .version        = NBD_GENL_VERSION,
2212         .module         = THIS_MODULE,
2213         .ops            = nbd_connect_genl_ops,
2214         .n_ops          = ARRAY_SIZE(nbd_connect_genl_ops),
2215         .maxattr        = NBD_ATTR_MAX,
2216         .policy = nbd_attr_policy,
2217         .mcgrps         = nbd_mcast_grps,
2218         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2219 };
2220
2221 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2222 {
2223         struct nlattr *dev_opt;
2224         u8 connected = 0;
2225         int ret;
2226
2227         /* This is a little racey, but for status it's ok.  The
2228          * reason we don't take a ref here is because we can't
2229          * take a ref in the index == -1 case as we would need
2230          * to put under the nbd_index_mutex, which could
2231          * deadlock if we are configured to remove ourselves
2232          * once we're disconnected.
2233          */
2234         if (refcount_read(&nbd->config_refs))
2235                 connected = 1;
2236         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2237         if (!dev_opt)
2238                 return -EMSGSIZE;
2239         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2240         if (ret)
2241                 return -EMSGSIZE;
2242         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2243                          connected);
2244         if (ret)
2245                 return -EMSGSIZE;
2246         nla_nest_end(reply, dev_opt);
2247         return 0;
2248 }
2249
2250 static int status_cb(int id, void *ptr, void *data)
2251 {
2252         struct nbd_device *nbd = ptr;
2253         return populate_nbd_status(nbd, (struct sk_buff *)data);
2254 }
2255
2256 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2257 {
2258         struct nlattr *dev_list;
2259         struct sk_buff *reply;
2260         void *reply_head;
2261         size_t msg_size;
2262         int index = -1;
2263         int ret = -ENOMEM;
2264
2265         if (info->attrs[NBD_ATTR_INDEX])
2266                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2267
2268         mutex_lock(&nbd_index_mutex);
2269
2270         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2271                                   nla_attr_size(sizeof(u8)));
2272         msg_size *= (index == -1) ? nbd_total_devices : 1;
2273
2274         reply = genlmsg_new(msg_size, GFP_KERNEL);
2275         if (!reply)
2276                 goto out;
2277         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2278                                        NBD_CMD_STATUS);
2279         if (!reply_head) {
2280                 nlmsg_free(reply);
2281                 goto out;
2282         }
2283
2284         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2285         if (index == -1) {
2286                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2287                 if (ret) {
2288                         nlmsg_free(reply);
2289                         goto out;
2290                 }
2291         } else {
2292                 struct nbd_device *nbd;
2293                 nbd = idr_find(&nbd_index_idr, index);
2294                 if (nbd) {
2295                         ret = populate_nbd_status(nbd, reply);
2296                         if (ret) {
2297                                 nlmsg_free(reply);
2298                                 goto out;
2299                         }
2300                 }
2301         }
2302         nla_nest_end(reply, dev_list);
2303         genlmsg_end(reply, reply_head);
2304         ret = genlmsg_reply(reply, info);
2305 out:
2306         mutex_unlock(&nbd_index_mutex);
2307         return ret;
2308 }
2309
2310 static void nbd_connect_reply(struct genl_info *info, int index)
2311 {
2312         struct sk_buff *skb;
2313         void *msg_head;
2314         int ret;
2315
2316         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2317         if (!skb)
2318                 return;
2319         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2320                                      NBD_CMD_CONNECT);
2321         if (!msg_head) {
2322                 nlmsg_free(skb);
2323                 return;
2324         }
2325         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2326         if (ret) {
2327                 nlmsg_free(skb);
2328                 return;
2329         }
2330         genlmsg_end(skb, msg_head);
2331         genlmsg_reply(skb, info);
2332 }
2333
2334 static void nbd_mcast_index(int index)
2335 {
2336         struct sk_buff *skb;
2337         void *msg_head;
2338         int ret;
2339
2340         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2341         if (!skb)
2342                 return;
2343         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2344                                      NBD_CMD_LINK_DEAD);
2345         if (!msg_head) {
2346                 nlmsg_free(skb);
2347                 return;
2348         }
2349         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2350         if (ret) {
2351                 nlmsg_free(skb);
2352                 return;
2353         }
2354         genlmsg_end(skb, msg_head);
2355         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2356 }
2357
2358 static void nbd_dead_link_work(struct work_struct *work)
2359 {
2360         struct link_dead_args *args = container_of(work, struct link_dead_args,
2361                                                    work);
2362         nbd_mcast_index(args->index);
2363         kfree(args);
2364 }
2365
2366 static int __init nbd_init(void)
2367 {
2368         int i;
2369
2370         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2371
2372         if (max_part < 0) {
2373                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2374                 return -EINVAL;
2375         }
2376
2377         part_shift = 0;
2378         if (max_part > 0) {
2379                 part_shift = fls(max_part);
2380
2381                 /*
2382                  * Adjust max_part according to part_shift as it is exported
2383                  * to user space so that user can know the max number of
2384                  * partition kernel should be able to manage.
2385                  *
2386                  * Note that -1 is required because partition 0 is reserved
2387                  * for the whole disk.
2388                  */
2389                 max_part = (1UL << part_shift) - 1;
2390         }
2391
2392         if ((1UL << part_shift) > DISK_MAX_PARTS)
2393                 return -EINVAL;
2394
2395         if (nbds_max > 1UL << (MINORBITS - part_shift))
2396                 return -EINVAL;
2397
2398         if (register_blkdev(NBD_MAJOR, "nbd"))
2399                 return -EIO;
2400
2401         if (genl_register_family(&nbd_genl_family)) {
2402                 unregister_blkdev(NBD_MAJOR, "nbd");
2403                 return -EINVAL;
2404         }
2405         nbd_dbg_init();
2406
2407         mutex_lock(&nbd_index_mutex);
2408         for (i = 0; i < nbds_max; i++)
2409                 nbd_dev_add(i);
2410         mutex_unlock(&nbd_index_mutex);
2411         return 0;
2412 }
2413
2414 static int nbd_exit_cb(int id, void *ptr, void *data)
2415 {
2416         struct list_head *list = (struct list_head *)data;
2417         struct nbd_device *nbd = ptr;
2418
2419         list_add_tail(&nbd->list, list);
2420         return 0;
2421 }
2422
2423 static void __exit nbd_cleanup(void)
2424 {
2425         struct nbd_device *nbd;
2426         LIST_HEAD(del_list);
2427
2428         nbd_dbg_close();
2429
2430         mutex_lock(&nbd_index_mutex);
2431         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2432         mutex_unlock(&nbd_index_mutex);
2433
2434         while (!list_empty(&del_list)) {
2435                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2436                 list_del_init(&nbd->list);
2437                 if (refcount_read(&nbd->refs) != 1)
2438                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2439                 nbd_put(nbd);
2440         }
2441
2442         idr_destroy(&nbd_index_idr);
2443         genl_unregister_family(&nbd_genl_family);
2444         unregister_blkdev(NBD_MAJOR, "nbd");
2445 }
2446
2447 module_init(nbd_init);
2448 module_exit(nbd_cleanup);
2449
2450 MODULE_DESCRIPTION("Network Block Device");
2451 MODULE_LICENSE("GPL");
2452
2453 module_param(nbds_max, int, 0444);
2454 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2455 module_param(max_part, int, 0444);
2456 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");