drivers/block/nbd.c

   1 /*
   2  * Network block device - make block devices work over TCP
   3  *
   4  * Note that you can not swap over this thing, yet. Seems to work but
   5  * deadlocks sometimes - you can not swap over TCP in general.
   6  *
   7  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   8  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
   9  *
  10  * This file is released under GPLv2 or later.
  11  *
  12  * (part of code stolen from loop.c)
  13  */
  14
  15 #include <linux/major.h>
  16
  17 #include <linux/blkdev.h>
  18 #include <linux/module.h>
  19 #include <linux/init.h>
  20 #include <linux/sched.h>
  21 #include <linux/sched/mm.h>
  22 #include <linux/fs.h>
  23 #include <linux/bio.h>
  24 #include <linux/stat.h>
  25 #include <linux/errno.h>
  26 #include <linux/file.h>
  27 #include <linux/ioctl.h>
  28 #include <linux/mutex.h>
  29 #include <linux/compiler.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 static DEFINE_IDR(nbd_index_idr);
  48 static DEFINE_MUTEX(nbd_index_mutex);
  49 static int nbd_total_devices = 0;
  50
  51 struct nbd_sock {
  52         struct socket *sock;
  53         struct mutex tx_lock;
  54         struct request *pending;
  55         int sent;
  56         bool dead;
  57         int fallback_index;
  58         int cookie;
  59 };
  60
  61 struct recv_thread_args {
  62         struct work_struct work;
  63         struct nbd_device *nbd;
  64         int index;
  65 };
  66
  67 struct link_dead_args {
  68         struct work_struct work;
  69         int index;
  70 };
  71
  72 #define NBD_TIMEDOUT                    0
  73 #define NBD_DISCONNECT_REQUESTED        1
  74 #define NBD_DISCONNECTED                2
  75 #define NBD_HAS_PID_FILE                3
  76 #define NBD_HAS_CONFIG_REF              4
  77 #define NBD_BOUND                       5
  78 #define NBD_DESTROY_ON_DISCONNECT       6
  79 #define NBD_DISCONNECT_ON_CLOSE         7
  80
  81 struct nbd_config {
  82         u32 flags;
  83         unsigned long runtime_flags;
  84         u64 dead_conn_timeout;
  85
  86         struct nbd_sock **socks;
  87         int num_connections;
  88         atomic_t live_connections;
  89         wait_queue_head_t conn_wait;
  90
  91         atomic_t recv_threads;
  92         wait_queue_head_t recv_wq;
  93         loff_t blksize;
  94         loff_t bytesize;
  95 #if IS_ENABLED(CONFIG_DEBUG_FS)
  96         struct dentry *dbg_dir;
  97 #endif
  98 };
  99
 100 struct nbd_device {
 101         struct blk_mq_tag_set tag_set;
 102
 103         int index;
 104         refcount_t config_refs;
 105         refcount_t refs;
 106         struct nbd_config *config;
 107         struct mutex config_lock;
 108         struct gendisk *disk;
 109
 110         struct list_head list;
 111         struct task_struct *task_recv;
 112         struct task_struct *task_setup;
 113 };
 114
 115 #define NBD_CMD_REQUEUED        1
 116
 117 struct nbd_cmd {
 118         struct nbd_device *nbd;
 119         struct mutex lock;
 120         int index;
 121         int cookie;
 122         blk_status_t status;
 123         unsigned long flags;
 124         u32 cmd_cookie;
 125 };
 126
 127 #if IS_ENABLED(CONFIG_DEBUG_FS)
 128 static struct dentry *nbd_dbg_dir;
 129 #endif
 130
 131 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 132
 133 #define NBD_MAGIC 0x68797548
 134
 135 static unsigned int nbds_max = 16;
 136 static int max_part = 16;
 137 static struct workqueue_struct *recv_workqueue;
 138 static int part_shift;
 139
 140 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 141 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 142 static void nbd_config_put(struct nbd_device *nbd);
 143 static void nbd_connect_reply(struct genl_info *info, int index);
 144 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 145 static void nbd_dead_link_work(struct work_struct *work);
 146 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 147
 148 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 149 {
 150         return disk_to_dev(nbd->disk);
 151 }
 152
 153 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 154 {
 155         struct request *req = blk_mq_rq_from_pdu(cmd);
 156
 157         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 158                 blk_mq_requeue_request(req, true);
 159 }
 160
 161 #define NBD_COOKIE_BITS 32
 162
 163 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 164 {
 165         struct request *req = blk_mq_rq_from_pdu(cmd);
 166         u32 tag = blk_mq_unique_tag(req);
 167         u64 cookie = cmd->cmd_cookie;
 168
 169         return (cookie << NBD_COOKIE_BITS) | tag;
 170 }
 171
 172 static u32 nbd_handle_to_tag(u64 handle)
 173 {
 174         return (u32)handle;
 175 }
 176
 177 static u32 nbd_handle_to_cookie(u64 handle)
 178 {
 179         return (u32)(handle >> NBD_COOKIE_BITS);
 180 }
 181
 182 static const char *nbdcmd_to_ascii(int cmd)
 183 {
 184         switch (cmd) {
 185         case  NBD_CMD_READ: return "read";
 186         case NBD_CMD_WRITE: return "write";
 187         case  NBD_CMD_DISC: return "disconnect";
 188         case NBD_CMD_FLUSH: return "flush";
 189         case  NBD_CMD_TRIM: return "trim/discard";
 190         }
 191         return "invalid";
 192 }
 193
 194 static ssize_t pid_show(struct device *dev,
 195                         struct device_attribute *attr, char *buf)
 196 {
 197         struct gendisk *disk = dev_to_disk(dev);
 198         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 199
 200         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 201 }
 202
 203 static const struct device_attribute pid_attr = {
 204         .attr = { .name = "pid", .mode = 0444},
 205         .show = pid_show,
 206 };
 207
 208 static void nbd_dev_remove(struct nbd_device *nbd)
 209 {
 210         struct gendisk *disk = nbd->disk;
 211         struct request_queue *q;
 212
 213         if (disk) {
 214                 q = disk->queue;
 215                 del_gendisk(disk);
 216                 blk_cleanup_queue(q);
 217                 blk_mq_free_tag_set(&nbd->tag_set);
 218                 disk->private_data = NULL;
 219                 put_disk(disk);
 220         }
 221         kfree(nbd);
 222 }
 223
 224 static void nbd_put(struct nbd_device *nbd)
 225 {
 226         if (refcount_dec_and_mutex_lock(&nbd->refs,
 227                                         &nbd_index_mutex)) {
 228                 idr_remove(&nbd_index_idr, nbd->index);
 229                 mutex_unlock(&nbd_index_mutex);
 230                 nbd_dev_remove(nbd);
 231         }
 232 }
 233
 234 static int nbd_disconnected(struct nbd_config *config)
 235 {
 236         return test_bit(NBD_DISCONNECTED, &config->runtime_flags) ||
 237                 test_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags);
 238 }
 239
 240 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 241                                 int notify)
 242 {
 243         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 244                 struct link_dead_args *args;
 245                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 246                 if (args) {
 247                         INIT_WORK(&args->work, nbd_dead_link_work);
 248                         args->index = nbd->index;
 249                         queue_work(system_wq, &args->work);
 250                 }
 251         }
 252         if (!nsock->dead) {
 253                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 254                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 255                         if (test_and_clear_bit(NBD_DISCONNECT_REQUESTED,
 256                                                &nbd->config->runtime_flags)) {
 257                                 set_bit(NBD_DISCONNECTED,
 258                                         &nbd->config->runtime_flags);
 259                                 dev_info(nbd_to_dev(nbd),
 260                                         "Disconnected due to user request.\n");
 261                         }
 262                 }
 263         }
 264         nsock->dead = true;
 265         nsock->pending = NULL;
 266         nsock->sent = 0;
 267 }
 268
 269 static void nbd_size_clear(struct nbd_device *nbd)
 270 {
 271         if (nbd->config->bytesize) {
 272                 set_capacity(nbd->disk, 0);
 273                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 274         }
 275 }
 276
 277 static void nbd_size_update(struct nbd_device *nbd)
 278 {
 279         struct nbd_config *config = nbd->config;
 280         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 281
 282         if (config->flags & NBD_FLAG_SEND_TRIM) {
 283                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 284                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 285                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 286         }
 287         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 288         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 289         set_capacity(nbd->disk, config->bytesize >> 9);
 290         if (bdev) {
 291                 if (bdev->bd_disk)
 292                         bd_set_size(bdev, config->bytesize);
 293                 else
 294                         bdev->bd_invalidated = 1;
 295                 bdput(bdev);
 296         }
 297         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 298 }
 299
 300 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 301                          loff_t nr_blocks)
 302 {
 303         struct nbd_config *config = nbd->config;
 304         config->blksize = blocksize;
 305         config->bytesize = blocksize * nr_blocks;
 306         if (nbd->task_recv != NULL)
 307                 nbd_size_update(nbd);
 308 }
 309
 310 static void nbd_complete_rq(struct request *req)
 311 {
 312         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 313
 314         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 315                 cmd->status ? "failed" : "done");
 316
 317         blk_mq_end_request(req, cmd->status);
 318 }
 319
 320 /*
 321  * Forcibly shutdown the socket causing all listeners to error
 322  */
 323 static void sock_shutdown(struct nbd_device *nbd)
 324 {
 325         struct nbd_config *config = nbd->config;
 326         int i;
 327
 328         if (config->num_connections == 0)
 329                 return;
 330         if (test_and_set_bit(NBD_DISCONNECTED, &config->runtime_flags))
 331                 return;
 332
 333         for (i = 0; i < config->num_connections; i++) {
 334                 struct nbd_sock *nsock = config->socks[i];
 335                 mutex_lock(&nsock->tx_lock);
 336                 nbd_mark_nsock_dead(nbd, nsock, 0);
 337                 mutex_unlock(&nsock->tx_lock);
 338         }
 339         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 340 }
 341
 342 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 343                                                  bool reserved)
 344 {
 345         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 346         struct nbd_device *nbd = cmd->nbd;
 347         struct nbd_config *config;
 348
 349         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 350                 cmd->status = BLK_STS_TIMEOUT;
 351                 goto done;
 352         }
 353         config = nbd->config;
 354
 355         if (!mutex_trylock(&cmd->lock))
 356                 return BLK_EH_RESET_TIMER;
 357
 358         if (config->num_connections > 1) {
 359                 dev_err_ratelimited(nbd_to_dev(nbd),
 360                                     "Connection timed out, retrying (%d/%d alive)\n",
 361                                     atomic_read(&config->live_connections),
 362                                     config->num_connections);
 363                 /*
 364                  * Hooray we have more connections, requeue this IO, the submit
 365                  * path will put it on a real connection.
 366                  */
 367                 if (config->socks && config->num_connections > 1) {
 368                         if (cmd->index < config->num_connections) {
 369                                 struct nbd_sock *nsock =
 370                                         config->socks[cmd->index];
 371                                 mutex_lock(&nsock->tx_lock);
 372                                 /* We can have multiple outstanding requests, so
 373                                  * we don't want to mark the nsock dead if we've
 374                                  * already reconnected with a new socket, so
 375                                  * only mark it dead if its the same socket we
 376                                  * were sent out on.
 377                                  */
 378                                 if (cmd->cookie == nsock->cookie)
 379                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 380                                 mutex_unlock(&nsock->tx_lock);
 381                         }
 382                         mutex_unlock(&cmd->lock);
 383                         nbd_requeue_cmd(cmd);
 384                         nbd_config_put(nbd);
 385                         return BLK_EH_DONE;
 386                 }
 387         } else {
 388                 dev_err_ratelimited(nbd_to_dev(nbd),
 389                                     "Connection timed out\n");
 390         }
 391         set_bit(NBD_TIMEDOUT, &config->runtime_flags);
 392         cmd->status = BLK_STS_IOERR;
 393         mutex_unlock(&cmd->lock);
 394         sock_shutdown(nbd);
 395         nbd_config_put(nbd);
 396 done:
 397         blk_mq_complete_request(req);
 398         return BLK_EH_DONE;
 399 }
 400
 401 /*
 402  *  Send or receive packet.
 403  */
 404 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 405                      struct iov_iter *iter, int msg_flags, int *sent)
 406 {
 407         struct nbd_config *config = nbd->config;
 408         struct socket *sock = config->socks[index]->sock;
 409         int result;
 410         struct msghdr msg;
 411         unsigned int noreclaim_flag;
 412
 413         if (unlikely(!sock)) {
 414                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 415                         "Attempted %s on closed socket in sock_xmit\n",
 416                         (send ? "send" : "recv"));
 417                 return -EINVAL;
 418         }
 419
 420         msg.msg_iter = *iter;
 421
 422         noreclaim_flag = memalloc_noreclaim_save();
 423         do {
 424                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 425                 msg.msg_name = NULL;
 426                 msg.msg_namelen = 0;
 427                 msg.msg_control = NULL;
 428                 msg.msg_controllen = 0;
 429                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 430
 431                 if (send)
 432                         result = sock_sendmsg(sock, &msg);
 433                 else
 434                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 435
 436                 if (result <= 0) {
 437                         if (result == 0)
 438                                 result = -EPIPE; /* short read */
 439                         break;
 440                 }
 441                 if (sent)
 442                         *sent += result;
 443         } while (msg_data_left(&msg));
 444
 445         memalloc_noreclaim_restore(noreclaim_flag);
 446
 447         return result;
 448 }
 449
 450 /*
 451  * Different settings for sk->sk_sndtimeo can result in different return values
 452  * if there is a signal pending when we enter sendmsg, because reasons?
 453  */
 454 static inline int was_interrupted(int result)
 455 {
 456         return result == -ERESTARTSYS || result == -EINTR;
 457 }
 458
 459 /* always call with the tx_lock held */
 460 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 461 {
 462         struct request *req = blk_mq_rq_from_pdu(cmd);
 463         struct nbd_config *config = nbd->config;
 464         struct nbd_sock *nsock = config->socks[index];
 465         int result;
 466         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 467         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 468         struct iov_iter from;
 469         unsigned long size = blk_rq_bytes(req);
 470         struct bio *bio;
 471         u64 handle;
 472         u32 type;
 473         u32 nbd_cmd_flags = 0;
 474         int sent = nsock->sent, skip = 0;
 475
 476         iov_iter_kvec(&from, WRITE | ITER_KVEC, &iov, 1, sizeof(request));
 477
 478         switch (req_op(req)) {
 479         case REQ_OP_DISCARD:
 480                 type = NBD_CMD_TRIM;
 481                 break;
 482         case REQ_OP_FLUSH:
 483                 type = NBD_CMD_FLUSH;
 484                 break;
 485         case REQ_OP_WRITE:
 486                 type = NBD_CMD_WRITE;
 487                 break;
 488         case REQ_OP_READ:
 489                 type = NBD_CMD_READ;
 490                 break;
 491         default:
 492                 return -EIO;
 493         }
 494
 495         if (rq_data_dir(req) == WRITE &&
 496             (config->flags & NBD_FLAG_READ_ONLY)) {
 497                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 498                                     "Write on read-only\n");
 499                 return -EIO;
 500         }
 501
 502         if (req->cmd_flags & REQ_FUA)
 503                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 504
 505         /* We did a partial send previously, and we at least sent the whole
 506          * request struct, so just go and send the rest of the pages in the
 507          * request.
 508          */
 509         if (sent) {
 510                 if (sent >= sizeof(request)) {
 511                         skip = sent - sizeof(request);
 512                         goto send_pages;
 513                 }
 514                 iov_iter_advance(&from, sent);
 515         } else {
 516                 cmd->cmd_cookie++;
 517         }
 518         cmd->index = index;
 519         cmd->cookie = nsock->cookie;
 520         request.type = htonl(type | nbd_cmd_flags);
 521         if (type != NBD_CMD_FLUSH) {
 522                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 523                 request.len = htonl(size);
 524         }
 525         handle = nbd_cmd_handle(cmd);
 526         memcpy(request.handle, &handle, sizeof(handle));
 527
 528         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 529                 req, nbdcmd_to_ascii(type),
 530                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 531         result = sock_xmit(nbd, index, 1, &from,
 532                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 533         if (result <= 0) {
 534                 if (was_interrupted(result)) {
 535                         /* If we havne't sent anything we can just return BUSY,
 536                          * however if we have sent something we need to make
 537                          * sure we only allow this req to be sent until we are
 538                          * completely done.
 539                          */
 540                         if (sent) {
 541                                 nsock->pending = req;
 542                                 nsock->sent = sent;
 543                         }
 544                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 545                         return BLK_STS_RESOURCE;
 546                 }
 547                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 548                         "Send control failed (result %d)\n", result);
 549                 return -EAGAIN;
 550         }
 551 send_pages:
 552         if (type != NBD_CMD_WRITE)
 553                 goto out;
 554
 555         bio = req->bio;
 556         while (bio) {
 557                 struct bio *next = bio->bi_next;
 558                 struct bvec_iter iter;
 559                 struct bio_vec bvec;
 560
 561                 bio_for_each_segment(bvec, bio, iter) {
 562                         bool is_last = !next && bio_iter_last(bvec, iter);
 563                         int flags = is_last ? 0 : MSG_MORE;
 564
 565                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 566                                 req, bvec.bv_len);
 567                         iov_iter_bvec(&from, ITER_BVEC | WRITE,
 568                                       &bvec, 1, bvec.bv_len);
 569                         if (skip) {
 570                                 if (skip >= iov_iter_count(&from)) {
 571                                         skip -= iov_iter_count(&from);
 572                                         continue;
 573                                 }
 574                                 iov_iter_advance(&from, skip);
 575                                 skip = 0;
 576                         }
 577                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 578                         if (result <= 0) {
 579                                 if (was_interrupted(result)) {
 580                                         /* We've already sent the header, we
 581                                          * have no choice but to set pending and
 582                                          * return BUSY.
 583                                          */
 584                                         nsock->pending = req;
 585                                         nsock->sent = sent;
 586                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 587                                         return BLK_STS_RESOURCE;
 588                                 }
 589                                 dev_err(disk_to_dev(nbd->disk),
 590                                         "Send data failed (result %d)\n",
 591                                         result);
 592                                 return -EAGAIN;
 593                         }
 594                         /*
 595                          * The completion might already have come in,
 596                          * so break for the last one instead of letting
 597                          * the iterator do it. This prevents use-after-free
 598                          * of the bio.
 599                          */
 600                         if (is_last)
 601                                 break;
 602                 }
 603                 bio = next;
 604         }
 605 out:
 606         nsock->pending = NULL;
 607         nsock->sent = 0;
 608         return 0;
 609 }
 610
 611 /* NULL returned = something went wrong, inform userspace */
 612 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 613 {
 614         struct nbd_config *config = nbd->config;
 615         int result;
 616         struct nbd_reply reply;
 617         struct nbd_cmd *cmd;
 618         struct request *req = NULL;
 619         u64 handle;
 620         u16 hwq;
 621         u32 tag;
 622         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 623         struct iov_iter to;
 624         int ret = 0;
 625
 626         reply.magic = 0;
 627         iov_iter_kvec(&to, READ | ITER_KVEC, &iov, 1, sizeof(reply));
 628         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 629         if (result <= 0) {
 630                 if (!nbd_disconnected(config))
 631                         dev_err(disk_to_dev(nbd->disk),
 632                                 "Receive control failed (result %d)\n", result);
 633                 return ERR_PTR(result);
 634         }
 635
 636         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 637                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 638                                 (unsigned long)ntohl(reply.magic));
 639                 return ERR_PTR(-EPROTO);
 640         }
 641
 642         memcpy(&handle, reply.handle, sizeof(handle));
 643         tag = nbd_handle_to_tag(handle);
 644         hwq = blk_mq_unique_tag_to_hwq(tag);
 645         if (hwq < nbd->tag_set.nr_hw_queues)
 646                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 647                                        blk_mq_unique_tag_to_tag(tag));
 648         if (!req || !blk_mq_request_started(req)) {
 649                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 650                         tag, req);
 651                 return ERR_PTR(-ENOENT);
 652         }
 653         cmd = blk_mq_rq_to_pdu(req);
 654
 655         mutex_lock(&cmd->lock);
 656         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 657                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 658                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 659                 ret = -ENOENT;
 660                 goto out;
 661         }
 662         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 663                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 664                         req);
 665                 ret = -ENOENT;
 666                 goto out;
 667         }
 668         if (ntohl(reply.error)) {
 669                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 670                         ntohl(reply.error));
 671                 cmd->status = BLK_STS_IOERR;
 672                 goto out;
 673         }
 674
 675         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 676         if (rq_data_dir(req) != WRITE) {
 677                 struct req_iterator iter;
 678                 struct bio_vec bvec;
 679
 680                 rq_for_each_segment(bvec, req, iter) {
 681                         iov_iter_bvec(&to, ITER_BVEC | READ,
 682                                       &bvec, 1, bvec.bv_len);
 683                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 684                         if (result <= 0) {
 685                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 686                                         result);
 687                                 /*
 688                                  * If we've disconnected or we only have 1
 689                                  * connection then we need to make sure we
 690                                  * complete this request, otherwise error out
 691                                  * and let the timeout stuff handle resubmitting
 692                                  * this request onto another connection.
 693                                  */
 694                                 if (nbd_disconnected(config) ||
 695                                     config->num_connections <= 1) {
 696                                         cmd->status = BLK_STS_IOERR;
 697                                         goto out;
 698                                 }
 699                                 ret = -EIO;
 700                                 goto out;
 701                         }
 702                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 703                                 req, bvec.bv_len);
 704                 }
 705         }
 706 out:
 707         mutex_unlock(&cmd->lock);
 708         return ret ? ERR_PTR(ret) : cmd;
 709 }
 710
 711 static void recv_work(struct work_struct *work)
 712 {
 713         struct recv_thread_args *args = container_of(work,
 714                                                      struct recv_thread_args,
 715                                                      work);
 716         struct nbd_device *nbd = args->nbd;
 717         struct nbd_config *config = nbd->config;
 718         struct nbd_cmd *cmd;
 719
 720         while (1) {
 721                 cmd = nbd_read_stat(nbd, args->index);
 722                 if (IS_ERR(cmd)) {
 723                         struct nbd_sock *nsock = config->socks[args->index];
 724
 725                         mutex_lock(&nsock->tx_lock);
 726                         nbd_mark_nsock_dead(nbd, nsock, 1);
 727                         mutex_unlock(&nsock->tx_lock);
 728                         break;
 729                 }
 730
 731                 blk_mq_complete_request(blk_mq_rq_from_pdu(cmd));
 732         }
 733         atomic_dec(&config->recv_threads);
 734         wake_up(&config->recv_wq);
 735         nbd_config_put(nbd);
 736         kfree(args);
 737 }
 738
 739 static void nbd_clear_req(struct request *req, void *data, bool reserved)
 740 {
 741         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 742
 743         cmd->status = BLK_STS_IOERR;
 744         blk_mq_complete_request(req);
 745 }
 746
 747 static void nbd_clear_que(struct nbd_device *nbd)
 748 {
 749         blk_mq_quiesce_queue(nbd->disk->queue);
 750         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 751         blk_mq_unquiesce_queue(nbd->disk->queue);
 752         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 753 }
 754
 755 static int find_fallback(struct nbd_device *nbd, int index)
 756 {
 757         struct nbd_config *config = nbd->config;
 758         int new_index = -1;
 759         struct nbd_sock *nsock = config->socks[index];
 760         int fallback = nsock->fallback_index;
 761
 762         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 763                 return new_index;
 764
 765         if (config->num_connections <= 1) {
 766                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 767                                     "Attempted send on invalid socket\n");
 768                 return new_index;
 769         }
 770
 771         if (fallback >= 0 && fallback < config->num_connections &&
 772             !config->socks[fallback]->dead)
 773                 return fallback;
 774
 775         if (nsock->fallback_index < 0 ||
 776             nsock->fallback_index >= config->num_connections ||
 777             config->socks[nsock->fallback_index]->dead) {
 778                 int i;
 779                 for (i = 0; i < config->num_connections; i++) {
 780                         if (i == index)
 781                                 continue;
 782                         if (!config->socks[i]->dead) {
 783                                 new_index = i;
 784                                 break;
 785                         }
 786                 }
 787                 nsock->fallback_index = new_index;
 788                 if (new_index < 0) {
 789                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 790                                             "Dead connection, failed to find a fallback\n");
 791                         return new_index;
 792                 }
 793         }
 794         new_index = nsock->fallback_index;
 795         return new_index;
 796 }
 797
 798 static int wait_for_reconnect(struct nbd_device *nbd)
 799 {
 800         struct nbd_config *config = nbd->config;
 801         if (!config->dead_conn_timeout)
 802                 return 0;
 803         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 804                 return 0;
 805         return wait_event_timeout(config->conn_wait,
 806                                   atomic_read(&config->live_connections) > 0,
 807                                   config->dead_conn_timeout) > 0;
 808 }
 809
 810 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 811 {
 812         struct request *req = blk_mq_rq_from_pdu(cmd);
 813         struct nbd_device *nbd = cmd->nbd;
 814         struct nbd_config *config;
 815         struct nbd_sock *nsock;
 816         int ret;
 817
 818         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 819                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 820                                     "Socks array is empty\n");
 821                 blk_mq_start_request(req);
 822                 return -EINVAL;
 823         }
 824         config = nbd->config;
 825
 826         if (index >= config->num_connections) {
 827                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 828                                     "Attempted send on invalid socket\n");
 829                 nbd_config_put(nbd);
 830                 blk_mq_start_request(req);
 831                 return -EINVAL;
 832         }
 833         cmd->status = BLK_STS_OK;
 834 again:
 835         nsock = config->socks[index];
 836         mutex_lock(&nsock->tx_lock);
 837         if (nsock->dead) {
 838                 int old_index = index;
 839                 index = find_fallback(nbd, index);
 840                 mutex_unlock(&nsock->tx_lock);
 841                 if (index < 0) {
 842                         if (wait_for_reconnect(nbd)) {
 843                                 index = old_index;
 844                                 goto again;
 845                         }
 846                         /* All the sockets should already be down at this point,
 847                          * we just want to make sure that DISCONNECTED is set so
 848                          * any requests that come in that were queue'ed waiting
 849                          * for the reconnect timer don't trigger the timer again
 850                          * and instead just error out.
 851                          */
 852                         sock_shutdown(nbd);
 853                         nbd_config_put(nbd);
 854                         blk_mq_start_request(req);
 855                         return -EIO;
 856                 }
 857                 goto again;
 858         }
 859
 860         /* Handle the case that we have a pending request that was partially
 861          * transmitted that _has_ to be serviced first.  We need to call requeue
 862          * here so that it gets put _after_ the request that is already on the
 863          * dispatch list.
 864          */
 865         blk_mq_start_request(req);
 866         if (unlikely(nsock->pending && nsock->pending != req)) {
 867                 nbd_requeue_cmd(cmd);
 868                 ret = 0;
 869                 goto out;
 870         }
 871         /*
 872          * Some failures are related to the link going down, so anything that
 873          * returns EAGAIN can be retried on a different socket.
 874          */
 875         ret = nbd_send_cmd(nbd, cmd, index);
 876         if (ret == -EAGAIN) {
 877                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 878                                     "Request send failed, requeueing\n");
 879                 nbd_mark_nsock_dead(nbd, nsock, 1);
 880                 nbd_requeue_cmd(cmd);
 881                 ret = 0;
 882         }
 883 out:
 884         mutex_unlock(&nsock->tx_lock);
 885         nbd_config_put(nbd);
 886         return ret;
 887 }
 888
 889 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 890                         const struct blk_mq_queue_data *bd)
 891 {
 892         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 893         int ret;
 894
 895         /*
 896          * Since we look at the bio's to send the request over the network we
 897          * need to make sure the completion work doesn't mark this request done
 898          * before we are done doing our send.  This keeps us from dereferencing
 899          * freed data if we have particularly fast completions (ie we get the
 900          * completion before we exit sock_xmit on the last bvec) or in the case
 901          * that the server is misbehaving (or there was an error) before we're
 902          * done sending everything over the wire.
 903          */
 904         mutex_lock(&cmd->lock);
 905         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 906
 907         /* We can be called directly from the user space process, which means we
 908          * could possibly have signals pending so our sendmsg will fail.  In
 909          * this case we need to return that we are busy, otherwise error out as
 910          * appropriate.
 911          */
 912         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 913         if (ret < 0)
 914                 ret = BLK_STS_IOERR;
 915         else if (!ret)
 916                 ret = BLK_STS_OK;
 917         mutex_unlock(&cmd->lock);
 918
 919         return ret;
 920 }
 921
 922 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
 923                           bool netlink)
 924 {
 925         struct nbd_config *config = nbd->config;
 926         struct socket *sock;
 927         struct nbd_sock **socks;
 928         struct nbd_sock *nsock;
 929         int err;
 930
 931         sock = sockfd_lookup(arg, &err);
 932         if (!sock)
 933                 return err;
 934
 935         if (!netlink && !nbd->task_setup &&
 936             !test_bit(NBD_BOUND, &config->runtime_flags))
 937                 nbd->task_setup = current;
 938
 939         if (!netlink &&
 940             (nbd->task_setup != current ||
 941              test_bit(NBD_BOUND, &config->runtime_flags))) {
 942                 dev_err(disk_to_dev(nbd->disk),
 943                         "Device being setup by another task");
 944                 sockfd_put(sock);
 945                 return -EBUSY;
 946         }
 947
 948         socks = krealloc(config->socks, (config->num_connections + 1) *
 949                          sizeof(struct nbd_sock *), GFP_KERNEL);
 950         if (!socks) {
 951                 sockfd_put(sock);
 952                 return -ENOMEM;
 953         }
 954         nsock = kzalloc(sizeof(struct nbd_sock), GFP_KERNEL);
 955         if (!nsock) {
 956                 sockfd_put(sock);
 957                 return -ENOMEM;
 958         }
 959
 960         config->socks = socks;
 961
 962         nsock->fallback_index = -1;
 963         nsock->dead = false;
 964         mutex_init(&nsock->tx_lock);
 965         nsock->sock = sock;
 966         nsock->pending = NULL;
 967         nsock->sent = 0;
 968         nsock->cookie = 0;
 969         socks[config->num_connections++] = nsock;
 970         atomic_inc(&config->live_connections);
 971
 972         return 0;
 973 }
 974
 975 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
 976 {
 977         struct nbd_config *config = nbd->config;
 978         struct socket *sock, *old;
 979         struct recv_thread_args *args;
 980         int i;
 981         int err;
 982
 983         sock = sockfd_lookup(arg, &err);
 984         if (!sock)
 985                 return err;
 986
 987         args = kzalloc(sizeof(*args), GFP_KERNEL);
 988         if (!args) {
 989                 sockfd_put(sock);
 990                 return -ENOMEM;
 991         }
 992
 993         for (i = 0; i < config->num_connections; i++) {
 994                 struct nbd_sock *nsock = config->socks[i];
 995
 996                 if (!nsock->dead)
 997                         continue;
 998
 999                 mutex_lock(&nsock->tx_lock);
1000                 if (!nsock->dead) {
1001                         mutex_unlock(&nsock->tx_lock);
1002                         continue;
1003                 }
1004                 sk_set_memalloc(sock->sk);
1005                 if (nbd->tag_set.timeout)
1006                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1007                 atomic_inc(&config->recv_threads);
1008                 refcount_inc(&nbd->config_refs);
1009                 old = nsock->sock;
1010                 nsock->fallback_index = -1;
1011                 nsock->sock = sock;
1012                 nsock->dead = false;
1013                 INIT_WORK(&args->work, recv_work);
1014                 args->index = i;
1015                 args->nbd = nbd;
1016                 nsock->cookie++;
1017                 mutex_unlock(&nsock->tx_lock);
1018                 sockfd_put(old);
1019
1020                 clear_bit(NBD_DISCONNECTED, &config->runtime_flags);
1021
1022                 /* We take the tx_mutex in an error path in the recv_work, so we
1023                  * need to queue_work outside of the tx_mutex.
1024                  */
1025                 queue_work(recv_workqueue, &args->work);
1026
1027                 atomic_inc(&config->live_connections);
1028                 wake_up(&config->conn_wait);
1029                 return 0;
1030         }
1031         sockfd_put(sock);
1032         kfree(args);
1033         return -ENOSPC;
1034 }
1035
1036 static void nbd_bdev_reset(struct block_device *bdev)
1037 {
1038         if (bdev->bd_openers > 1)
1039                 return;
1040         bd_set_size(bdev, 0);
1041 }
1042
1043 static void nbd_parse_flags(struct nbd_device *nbd)
1044 {
1045         struct nbd_config *config = nbd->config;
1046         if (config->flags & NBD_FLAG_READ_ONLY)
1047                 set_disk_ro(nbd->disk, true);
1048         else
1049                 set_disk_ro(nbd->disk, false);
1050         if (config->flags & NBD_FLAG_SEND_TRIM)
1051                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1052         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1053                 if (config->flags & NBD_FLAG_SEND_FUA)
1054                         blk_queue_write_cache(nbd->disk->queue, true, true);
1055                 else
1056                         blk_queue_write_cache(nbd->disk->queue, true, false);
1057         }
1058         else
1059                 blk_queue_write_cache(nbd->disk->queue, false, false);
1060 }
1061
1062 static void send_disconnects(struct nbd_device *nbd)
1063 {
1064         struct nbd_config *config = nbd->config;
1065         struct nbd_request request = {
1066                 .magic = htonl(NBD_REQUEST_MAGIC),
1067                 .type = htonl(NBD_CMD_DISC),
1068         };
1069         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1070         struct iov_iter from;
1071         int i, ret;
1072
1073         for (i = 0; i < config->num_connections; i++) {
1074                 struct nbd_sock *nsock = config->socks[i];
1075
1076                 iov_iter_kvec(&from, WRITE | ITER_KVEC, &iov, 1, sizeof(request));
1077                 mutex_lock(&nsock->tx_lock);
1078                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1079                 if (ret <= 0)
1080                         dev_err(disk_to_dev(nbd->disk),
1081                                 "Send disconnect failed %d\n", ret);
1082                 mutex_unlock(&nsock->tx_lock);
1083         }
1084 }
1085
1086 static int nbd_disconnect(struct nbd_device *nbd)
1087 {
1088         struct nbd_config *config = nbd->config;
1089
1090         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1091         set_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags);
1092         send_disconnects(nbd);
1093         return 0;
1094 }
1095
1096 static void nbd_clear_sock(struct nbd_device *nbd)
1097 {
1098         sock_shutdown(nbd);
1099         nbd_clear_que(nbd);
1100         nbd->task_setup = NULL;
1101 }
1102
1103 static void nbd_config_put(struct nbd_device *nbd)
1104 {
1105         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1106                                         &nbd->config_lock)) {
1107                 struct nbd_config *config = nbd->config;
1108                 nbd_dev_dbg_close(nbd);
1109                 nbd_size_clear(nbd);
1110                 if (test_and_clear_bit(NBD_HAS_PID_FILE,
1111                                        &config->runtime_flags))
1112                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1113                 nbd->task_recv = NULL;
1114                 nbd_clear_sock(nbd);
1115                 if (config->num_connections) {
1116                         int i;
1117                         for (i = 0; i < config->num_connections; i++) {
1118                                 sockfd_put(config->socks[i]->sock);
1119                                 kfree(config->socks[i]);
1120                         }
1121                         kfree(config->socks);
1122                 }
1123                 kfree(nbd->config);
1124                 nbd->config = NULL;
1125
1126                 nbd->tag_set.timeout = 0;
1127                 nbd->disk->queue->limits.discard_granularity = 0;
1128                 nbd->disk->queue->limits.discard_alignment = 0;
1129                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1130                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1131
1132                 mutex_unlock(&nbd->config_lock);
1133                 nbd_put(nbd);
1134                 module_put(THIS_MODULE);
1135         }
1136 }
1137
1138 static int nbd_start_device(struct nbd_device *nbd)
1139 {
1140         struct nbd_config *config = nbd->config;
1141         int num_connections = config->num_connections;
1142         int error = 0, i;
1143
1144         if (nbd->task_recv)
1145                 return -EBUSY;
1146         if (!config->socks)
1147                 return -EINVAL;
1148         if (num_connections > 1 &&
1149             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1150                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1151                 return -EINVAL;
1152         }
1153
1154         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1155         nbd->task_recv = current;
1156
1157         nbd_parse_flags(nbd);
1158
1159         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1160         if (error) {
1161                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1162                 return error;
1163         }
1164         set_bit(NBD_HAS_PID_FILE, &config->runtime_flags);
1165
1166         nbd_dev_dbg_init(nbd);
1167         for (i = 0; i < num_connections; i++) {
1168                 struct recv_thread_args *args;
1169
1170                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1171                 if (!args) {
1172                         sock_shutdown(nbd);
1173                         return -ENOMEM;
1174                 }
1175                 sk_set_memalloc(config->socks[i]->sock->sk);
1176                 if (nbd->tag_set.timeout)
1177                         config->socks[i]->sock->sk->sk_sndtimeo =
1178                                 nbd->tag_set.timeout;
1179                 atomic_inc(&config->recv_threads);
1180                 refcount_inc(&nbd->config_refs);
1181                 INIT_WORK(&args->work, recv_work);
1182                 args->nbd = nbd;
1183                 args->index = i;
1184                 queue_work(recv_workqueue, &args->work);
1185         }
1186         nbd_size_update(nbd);
1187         return error;
1188 }
1189
1190 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1191 {
1192         struct nbd_config *config = nbd->config;
1193         int ret;
1194
1195         ret = nbd_start_device(nbd);
1196         if (ret)
1197                 return ret;
1198
1199         if (max_part)
1200                 bdev->bd_invalidated = 1;
1201         mutex_unlock(&nbd->config_lock);
1202         ret = wait_event_interruptible(config->recv_wq,
1203                                          atomic_read(&config->recv_threads) == 0);
1204         if (ret)
1205                 sock_shutdown(nbd);
1206         mutex_lock(&nbd->config_lock);
1207         nbd_bdev_reset(bdev);
1208         /* user requested, ignore socket errors */
1209         if (test_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags))
1210                 ret = 0;
1211         if (test_bit(NBD_TIMEDOUT, &config->runtime_flags))
1212                 ret = -ETIMEDOUT;
1213         return ret;
1214 }
1215
1216 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1217                                  struct block_device *bdev)
1218 {
1219         sock_shutdown(nbd);
1220         kill_bdev(bdev);
1221         nbd_bdev_reset(bdev);
1222         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1223                                &nbd->config->runtime_flags))
1224                 nbd_config_put(nbd);
1225 }
1226
1227 /* Must be called with config_lock held */
1228 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1229                        unsigned int cmd, unsigned long arg)
1230 {
1231         struct nbd_config *config = nbd->config;
1232
1233         switch (cmd) {
1234         case NBD_DISCONNECT:
1235                 return nbd_disconnect(nbd);
1236         case NBD_CLEAR_SOCK:
1237                 nbd_clear_sock_ioctl(nbd, bdev);
1238                 return 0;
1239         case NBD_SET_SOCK:
1240                 return nbd_add_socket(nbd, arg, false);
1241         case NBD_SET_BLKSIZE:
1242                 if (!arg || !is_power_of_2(arg) || arg < 512 ||
1243                     arg > PAGE_SIZE)
1244                         return -EINVAL;
1245                 nbd_size_set(nbd, arg,
1246                              div_s64(config->bytesize, arg));
1247                 return 0;
1248         case NBD_SET_SIZE:
1249                 nbd_size_set(nbd, config->blksize,
1250                              div_s64(arg, config->blksize));
1251                 return 0;
1252         case NBD_SET_SIZE_BLOCKS:
1253                 nbd_size_set(nbd, config->blksize, arg);
1254                 return 0;
1255         case NBD_SET_TIMEOUT:
1256                 if (arg) {
1257                         nbd->tag_set.timeout = arg * HZ;
1258                         blk_queue_rq_timeout(nbd->disk->queue, arg * HZ);
1259                 }
1260                 return 0;
1261
1262         case NBD_SET_FLAGS:
1263                 config->flags = arg;
1264                 return 0;
1265         case NBD_DO_IT:
1266                 return nbd_start_device_ioctl(nbd, bdev);
1267         case NBD_CLEAR_QUE:
1268                 /*
1269                  * This is for compatibility only.  The queue is always cleared
1270                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1271                  */
1272                 return 0;
1273         case NBD_PRINT_DEBUG:
1274                 /*
1275                  * For compatibility only, we no longer keep a list of
1276                  * outstanding requests.
1277                  */
1278                 return 0;
1279         }
1280         return -ENOTTY;
1281 }
1282
1283 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1284                      unsigned int cmd, unsigned long arg)
1285 {
1286         struct nbd_device *nbd = bdev->bd_disk->private_data;
1287         struct nbd_config *config = nbd->config;
1288         int error = -EINVAL;
1289
1290         if (!capable(CAP_SYS_ADMIN))
1291                 return -EPERM;
1292
1293         /* The block layer will pass back some non-nbd ioctls in case we have
1294          * special handling for them, but we don't so just return an error.
1295          */
1296         if (_IOC_TYPE(cmd) != 0xab)
1297                 return -EINVAL;
1298
1299         mutex_lock(&nbd->config_lock);
1300
1301         /* Don't allow ioctl operations on a nbd device that was created with
1302          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1303          */
1304         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1305             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1306                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1307         else
1308                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1309         mutex_unlock(&nbd->config_lock);
1310         return error;
1311 }
1312
1313 static struct nbd_config *nbd_alloc_config(void)
1314 {
1315         struct nbd_config *config;
1316
1317         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1318         if (!config)
1319                 return NULL;
1320         atomic_set(&config->recv_threads, 0);
1321         init_waitqueue_head(&config->recv_wq);
1322         init_waitqueue_head(&config->conn_wait);
1323         config->blksize = 1024;
1324         atomic_set(&config->live_connections, 0);
1325         try_module_get(THIS_MODULE);
1326         return config;
1327 }
1328
1329 static int nbd_open(struct block_device *bdev, fmode_t mode)
1330 {
1331         struct nbd_device *nbd;
1332         int ret = 0;
1333
1334         mutex_lock(&nbd_index_mutex);
1335         nbd = bdev->bd_disk->private_data;
1336         if (!nbd) {
1337                 ret = -ENXIO;
1338                 goto out;
1339         }
1340         if (!refcount_inc_not_zero(&nbd->refs)) {
1341                 ret = -ENXIO;
1342                 goto out;
1343         }
1344         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1345                 struct nbd_config *config;
1346
1347                 mutex_lock(&nbd->config_lock);
1348                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1349                         mutex_unlock(&nbd->config_lock);
1350                         goto out;
1351                 }
1352                 config = nbd->config = nbd_alloc_config();
1353                 if (!config) {
1354                         ret = -ENOMEM;
1355                         mutex_unlock(&nbd->config_lock);
1356                         goto out;
1357                 }
1358                 refcount_set(&nbd->config_refs, 1);
1359                 refcount_inc(&nbd->refs);
1360                 mutex_unlock(&nbd->config_lock);
1361                 bdev->bd_invalidated = 1;
1362         } else if (nbd_disconnected(nbd->config)) {
1363                 bdev->bd_invalidated = 1;
1364         }
1365 out:
1366         mutex_unlock(&nbd_index_mutex);
1367         return ret;
1368 }
1369
1370 static void nbd_release(struct gendisk *disk, fmode_t mode)
1371 {
1372         struct nbd_device *nbd = disk->private_data;
1373         struct block_device *bdev = bdget_disk(disk, 0);
1374
1375         if (test_bit(NBD_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1376                         bdev->bd_openers == 0)
1377                 nbd_disconnect_and_put(nbd);
1378
1379         nbd_config_put(nbd);
1380         nbd_put(nbd);
1381 }
1382
1383 static const struct block_device_operations nbd_fops =
1384 {
1385         .owner =        THIS_MODULE,
1386         .open =         nbd_open,
1387         .release =      nbd_release,
1388         .ioctl =        nbd_ioctl,
1389         .compat_ioctl = nbd_ioctl,
1390 };
1391
1392 #if IS_ENABLED(CONFIG_DEBUG_FS)
1393
1394 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1395 {
1396         struct nbd_device *nbd = s->private;
1397
1398         if (nbd->task_recv)
1399                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1400
1401         return 0;
1402 }
1403
1404 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1405 {
1406         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1407 }
1408
1409 static const struct file_operations nbd_dbg_tasks_ops = {
1410         .open = nbd_dbg_tasks_open,
1411         .read = seq_read,
1412         .llseek = seq_lseek,
1413         .release = single_release,
1414 };
1415
1416 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1417 {
1418         struct nbd_device *nbd = s->private;
1419         u32 flags = nbd->config->flags;
1420
1421         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1422
1423         seq_puts(s, "Known flags:\n");
1424
1425         if (flags & NBD_FLAG_HAS_FLAGS)
1426                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1427         if (flags & NBD_FLAG_READ_ONLY)
1428                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1429         if (flags & NBD_FLAG_SEND_FLUSH)
1430                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1431         if (flags & NBD_FLAG_SEND_FUA)
1432                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1433         if (flags & NBD_FLAG_SEND_TRIM)
1434                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1435
1436         return 0;
1437 }
1438
1439 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1440 {
1441         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1442 }
1443
1444 static const struct file_operations nbd_dbg_flags_ops = {
1445         .open = nbd_dbg_flags_open,
1446         .read = seq_read,
1447         .llseek = seq_lseek,
1448         .release = single_release,
1449 };
1450
1451 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1452 {
1453         struct dentry *dir;
1454         struct nbd_config *config = nbd->config;
1455
1456         if (!nbd_dbg_dir)
1457                 return -EIO;
1458
1459         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1460         if (!dir) {
1461                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1462                         nbd_name(nbd));
1463                 return -EIO;
1464         }
1465         config->dbg_dir = dir;
1466
1467         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1468         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1469         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1470         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1471         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1472
1473         return 0;
1474 }
1475
1476 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1477 {
1478         debugfs_remove_recursive(nbd->config->dbg_dir);
1479 }
1480
1481 static int nbd_dbg_init(void)
1482 {
1483         struct dentry *dbg_dir;
1484
1485         dbg_dir = debugfs_create_dir("nbd", NULL);
1486         if (!dbg_dir)
1487                 return -EIO;
1488
1489         nbd_dbg_dir = dbg_dir;
1490
1491         return 0;
1492 }
1493
1494 static void nbd_dbg_close(void)
1495 {
1496         debugfs_remove_recursive(nbd_dbg_dir);
1497 }
1498
1499 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1500
1501 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1502 {
1503         return 0;
1504 }
1505
1506 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1507 {
1508 }
1509
1510 static int nbd_dbg_init(void)
1511 {
1512         return 0;
1513 }
1514
1515 static void nbd_dbg_close(void)
1516 {
1517 }
1518
1519 #endif
1520
1521 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1522                             unsigned int hctx_idx, unsigned int numa_node)
1523 {
1524         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1525         cmd->nbd = set->driver_data;
1526         cmd->flags = 0;
1527         mutex_init(&cmd->lock);
1528         return 0;
1529 }
1530
1531 static const struct blk_mq_ops nbd_mq_ops = {
1532         .queue_rq       = nbd_queue_rq,
1533         .complete       = nbd_complete_rq,
1534         .init_request   = nbd_init_request,
1535         .timeout        = nbd_xmit_timeout,
1536 };
1537
1538 static int nbd_dev_add(int index)
1539 {
1540         struct nbd_device *nbd;
1541         struct gendisk *disk;
1542         struct request_queue *q;
1543         int err = -ENOMEM;
1544
1545         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1546         if (!nbd)
1547                 goto out;
1548
1549         disk = alloc_disk(1 << part_shift);
1550         if (!disk)
1551                 goto out_free_nbd;
1552
1553         if (index >= 0) {
1554                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1555                                 GFP_KERNEL);
1556                 if (err == -ENOSPC)
1557                         err = -EEXIST;
1558         } else {
1559                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1560                 if (err >= 0)
1561                         index = err;
1562         }
1563         if (err < 0)
1564                 goto out_free_disk;
1565
1566         nbd->index = index;
1567         nbd->disk = disk;
1568         nbd->tag_set.ops = &nbd_mq_ops;
1569         nbd->tag_set.nr_hw_queues = 1;
1570         nbd->tag_set.queue_depth = 128;
1571         nbd->tag_set.numa_node = NUMA_NO_NODE;
1572         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1573         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1574                 BLK_MQ_F_SG_MERGE | BLK_MQ_F_BLOCKING;
1575         nbd->tag_set.driver_data = nbd;
1576
1577         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1578         if (err)
1579                 goto out_free_idr;
1580
1581         q = blk_mq_init_queue(&nbd->tag_set);
1582         if (IS_ERR(q)) {
1583                 err = PTR_ERR(q);
1584                 goto out_free_tags;
1585         }
1586         disk->queue = q;
1587
1588         /*
1589          * Tell the block layer that we are not a rotational device
1590          */
1591         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1592         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1593         disk->queue->limits.discard_granularity = 0;
1594         disk->queue->limits.discard_alignment = 0;
1595         blk_queue_max_discard_sectors(disk->queue, 0);
1596         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1597         blk_queue_max_segments(disk->queue, USHRT_MAX);
1598         blk_queue_max_hw_sectors(disk->queue, 65536);
1599         disk->queue->limits.max_sectors = 256;
1600
1601         mutex_init(&nbd->config_lock);
1602         refcount_set(&nbd->config_refs, 0);
1603         refcount_set(&nbd->refs, 1);
1604         INIT_LIST_HEAD(&nbd->list);
1605         disk->major = NBD_MAJOR;
1606         disk->first_minor = index << part_shift;
1607         disk->fops = &nbd_fops;
1608         disk->private_data = nbd;
1609         sprintf(disk->disk_name, "nbd%d", index);
1610         add_disk(disk);
1611         nbd_total_devices++;
1612         return index;
1613
1614 out_free_tags:
1615         blk_mq_free_tag_set(&nbd->tag_set);
1616 out_free_idr:
1617         idr_remove(&nbd_index_idr, index);
1618 out_free_disk:
1619         put_disk(disk);
1620 out_free_nbd:
1621         kfree(nbd);
1622 out:
1623         return err;
1624 }
1625
1626 static int find_free_cb(int id, void *ptr, void *data)
1627 {
1628         struct nbd_device *nbd = ptr;
1629         struct nbd_device **found = data;
1630
1631         if (!refcount_read(&nbd->config_refs)) {
1632                 *found = nbd;
1633                 return 1;
1634         }
1635         return 0;
1636 }
1637
1638 /* Netlink interface. */
1639 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1640         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1641         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1642         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1643         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1644         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1645         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1646         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1647         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1648         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1649 };
1650
1651 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1652         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1653 };
1654
1655 /* We don't use this right now since we don't parse the incoming list, but we
1656  * still want it here so userspace knows what to expect.
1657  */
1658 static const struct nla_policy __attribute__((unused))
1659 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1660         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1661         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1662 };
1663
1664 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1665 {
1666         struct nbd_device *nbd = NULL;
1667         struct nbd_config *config;
1668         int index = -1;
1669         int ret;
1670         bool put_dev = false;
1671
1672         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1673                 return -EPERM;
1674
1675         if (info->attrs[NBD_ATTR_INDEX])
1676                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1677         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1678                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1679                 return -EINVAL;
1680         }
1681         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1682                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1683                 return -EINVAL;
1684         }
1685 again:
1686         mutex_lock(&nbd_index_mutex);
1687         if (index == -1) {
1688                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1689                 if (ret == 0) {
1690                         int new_index;
1691                         new_index = nbd_dev_add(-1);
1692                         if (new_index < 0) {
1693                                 mutex_unlock(&nbd_index_mutex);
1694                                 printk(KERN_ERR "nbd: failed to add new device\n");
1695                                 return new_index;
1696                         }
1697                         nbd = idr_find(&nbd_index_idr, new_index);
1698                 }
1699         } else {
1700                 nbd = idr_find(&nbd_index_idr, index);
1701                 if (!nbd) {
1702                         ret = nbd_dev_add(index);
1703                         if (ret < 0) {
1704                                 mutex_unlock(&nbd_index_mutex);
1705                                 printk(KERN_ERR "nbd: failed to add new device\n");
1706                                 return ret;
1707                         }
1708                         nbd = idr_find(&nbd_index_idr, index);
1709                 }
1710         }
1711         if (!nbd) {
1712                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1713                        index);
1714                 mutex_unlock(&nbd_index_mutex);
1715                 return -EINVAL;
1716         }
1717         if (!refcount_inc_not_zero(&nbd->refs)) {
1718                 mutex_unlock(&nbd_index_mutex);
1719                 if (index == -1)
1720                         goto again;
1721                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1722                        index);
1723                 return -EINVAL;
1724         }
1725         mutex_unlock(&nbd_index_mutex);
1726
1727         mutex_lock(&nbd->config_lock);
1728         if (refcount_read(&nbd->config_refs)) {
1729                 mutex_unlock(&nbd->config_lock);
1730                 nbd_put(nbd);
1731                 if (index == -1)
1732                         goto again;
1733                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1734                 return -EBUSY;
1735         }
1736         if (WARN_ON(nbd->config)) {
1737                 mutex_unlock(&nbd->config_lock);
1738                 nbd_put(nbd);
1739                 return -EINVAL;
1740         }
1741         config = nbd->config = nbd_alloc_config();
1742         if (!nbd->config) {
1743                 mutex_unlock(&nbd->config_lock);
1744                 nbd_put(nbd);
1745                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1746                 return -ENOMEM;
1747         }
1748         refcount_set(&nbd->config_refs, 1);
1749         set_bit(NBD_BOUND, &config->runtime_flags);
1750
1751         if (info->attrs[NBD_ATTR_SIZE_BYTES]) {
1752                 u64 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1753                 nbd_size_set(nbd, config->blksize,
1754                              div64_u64(bytes, config->blksize));
1755         }
1756         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1757                 u64 bsize =
1758                         nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1759                 nbd_size_set(nbd, bsize, div64_u64(config->bytesize, bsize));
1760         }
1761         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1762                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1763                 nbd->tag_set.timeout = timeout * HZ;
1764                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1765         }
1766         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1767                 config->dead_conn_timeout =
1768                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1769                 config->dead_conn_timeout *= HZ;
1770         }
1771         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1772                 config->flags =
1773                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1774         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1775                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1776                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1777                         set_bit(NBD_DESTROY_ON_DISCONNECT,
1778                                 &config->runtime_flags);
1779                         put_dev = true;
1780                 }
1781                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1782                         set_bit(NBD_DISCONNECT_ON_CLOSE,
1783                                 &config->runtime_flags);
1784                 }
1785         }
1786
1787         if (info->attrs[NBD_ATTR_SOCKETS]) {
1788                 struct nlattr *attr;
1789                 int rem, fd;
1790
1791                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1792                                     rem) {
1793                         struct nlattr *socks[NBD_SOCK_MAX+1];
1794
1795                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1796                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1797                                 ret = -EINVAL;
1798                                 goto out;
1799                         }
1800                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1801                                                nbd_sock_policy, info->extack);
1802                         if (ret != 0) {
1803                                 printk(KERN_ERR "nbd: error processing sock list\n");
1804                                 ret = -EINVAL;
1805                                 goto out;
1806                         }
1807                         if (!socks[NBD_SOCK_FD])
1808                                 continue;
1809                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1810                         ret = nbd_add_socket(nbd, fd, true);
1811                         if (ret)
1812                                 goto out;
1813                 }
1814         }
1815         ret = nbd_start_device(nbd);
1816 out:
1817         mutex_unlock(&nbd->config_lock);
1818         if (!ret) {
1819                 set_bit(NBD_HAS_CONFIG_REF, &config->runtime_flags);
1820                 refcount_inc(&nbd->config_refs);
1821                 nbd_connect_reply(info, nbd->index);
1822         }
1823         nbd_config_put(nbd);
1824         if (put_dev)
1825                 nbd_put(nbd);
1826         return ret;
1827 }
1828
1829 static void nbd_disconnect_and_put(struct nbd_device *nbd)
1830 {
1831         mutex_lock(&nbd->config_lock);
1832         nbd_disconnect(nbd);
1833         nbd_clear_sock(nbd);
1834         mutex_unlock(&nbd->config_lock);
1835         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1836                                &nbd->config->runtime_flags))
1837                 nbd_config_put(nbd);
1838 }
1839
1840 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
1841 {
1842         struct nbd_device *nbd;
1843         int index;
1844
1845         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1846                 return -EPERM;
1847
1848         if (!info->attrs[NBD_ATTR_INDEX]) {
1849                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
1850                 return -EINVAL;
1851         }
1852         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1853         mutex_lock(&nbd_index_mutex);
1854         nbd = idr_find(&nbd_index_idr, index);
1855         if (!nbd) {
1856                 mutex_unlock(&nbd_index_mutex);
1857                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1858                        index);
1859                 return -EINVAL;
1860         }
1861         if (!refcount_inc_not_zero(&nbd->refs)) {
1862                 mutex_unlock(&nbd_index_mutex);
1863                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1864                        index);
1865                 return -EINVAL;
1866         }
1867         mutex_unlock(&nbd_index_mutex);
1868         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1869                 nbd_put(nbd);
1870                 return 0;
1871         }
1872         nbd_disconnect_and_put(nbd);
1873         nbd_config_put(nbd);
1874         nbd_put(nbd);
1875         return 0;
1876 }
1877
1878 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
1879 {
1880         struct nbd_device *nbd = NULL;
1881         struct nbd_config *config;
1882         int index;
1883         int ret = 0;
1884         bool put_dev = false;
1885
1886         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1887                 return -EPERM;
1888
1889         if (!info->attrs[NBD_ATTR_INDEX]) {
1890                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
1891                 return -EINVAL;
1892         }
1893         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1894         mutex_lock(&nbd_index_mutex);
1895         nbd = idr_find(&nbd_index_idr, index);
1896         if (!nbd) {
1897                 mutex_unlock(&nbd_index_mutex);
1898                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
1899                        index);
1900                 return -EINVAL;
1901         }
1902         if (!refcount_inc_not_zero(&nbd->refs)) {
1903                 mutex_unlock(&nbd_index_mutex);
1904                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1905                        index);
1906                 return -EINVAL;
1907         }
1908         mutex_unlock(&nbd_index_mutex);
1909
1910         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1911                 dev_err(nbd_to_dev(nbd),
1912                         "not configured, cannot reconfigure\n");
1913                 nbd_put(nbd);
1914                 return -EINVAL;
1915         }
1916
1917         mutex_lock(&nbd->config_lock);
1918         config = nbd->config;
1919         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1920             !nbd->task_recv) {
1921                 dev_err(nbd_to_dev(nbd),
1922                         "not configured, cannot reconfigure\n");
1923                 ret = -EINVAL;
1924                 goto out;
1925         }
1926
1927         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1928                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1929                 nbd->tag_set.timeout = timeout * HZ;
1930                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1931         }
1932         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1933                 config->dead_conn_timeout =
1934                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1935                 config->dead_conn_timeout *= HZ;
1936         }
1937         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1938                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1939                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1940                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1941                                               &config->runtime_flags))
1942                                 put_dev = true;
1943                 } else {
1944                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1945                                                &config->runtime_flags))
1946                                 refcount_inc(&nbd->refs);
1947                 }
1948
1949                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1950                         set_bit(NBD_DISCONNECT_ON_CLOSE,
1951                                         &config->runtime_flags);
1952                 } else {
1953                         clear_bit(NBD_DISCONNECT_ON_CLOSE,
1954                                         &config->runtime_flags);
1955                 }
1956         }
1957
1958         if (info->attrs[NBD_ATTR_SOCKETS]) {
1959                 struct nlattr *attr;
1960                 int rem, fd;
1961
1962                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1963                                     rem) {
1964                         struct nlattr *socks[NBD_SOCK_MAX+1];
1965
1966                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1967                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1968                                 ret = -EINVAL;
1969                                 goto out;
1970                         }
1971                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1972                                                nbd_sock_policy, info->extack);
1973                         if (ret != 0) {
1974                                 printk(KERN_ERR "nbd: error processing sock list\n");
1975                                 ret = -EINVAL;
1976                                 goto out;
1977                         }
1978                         if (!socks[NBD_SOCK_FD])
1979                                 continue;
1980                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1981                         ret = nbd_reconnect_socket(nbd, fd);
1982                         if (ret) {
1983                                 if (ret == -ENOSPC)
1984                                         ret = 0;
1985                                 goto out;
1986                         }
1987                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
1988                 }
1989         }
1990 out:
1991         mutex_unlock(&nbd->config_lock);
1992         nbd_config_put(nbd);
1993         nbd_put(nbd);
1994         if (put_dev)
1995                 nbd_put(nbd);
1996         return ret;
1997 }
1998
1999 static const struct genl_ops nbd_connect_genl_ops[] = {
2000         {
2001                 .cmd    = NBD_CMD_CONNECT,
2002                 .policy = nbd_attr_policy,
2003                 .doit   = nbd_genl_connect,
2004         },
2005         {
2006                 .cmd    = NBD_CMD_DISCONNECT,
2007                 .policy = nbd_attr_policy,
2008                 .doit   = nbd_genl_disconnect,
2009         },
2010         {
2011                 .cmd    = NBD_CMD_RECONFIGURE,
2012                 .policy = nbd_attr_policy,
2013                 .doit   = nbd_genl_reconfigure,
2014         },
2015         {
2016                 .cmd    = NBD_CMD_STATUS,
2017                 .policy = nbd_attr_policy,
2018                 .doit   = nbd_genl_status,
2019         },
2020 };
2021
2022 static const struct genl_multicast_group nbd_mcast_grps[] = {
2023         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2024 };
2025
2026 static struct genl_family nbd_genl_family __ro_after_init = {
2027         .hdrsize        = 0,
2028         .name           = NBD_GENL_FAMILY_NAME,
2029         .version        = NBD_GENL_VERSION,
2030         .module         = THIS_MODULE,
2031         .ops            = nbd_connect_genl_ops,
2032         .n_ops          = ARRAY_SIZE(nbd_connect_genl_ops),
2033         .maxattr        = NBD_ATTR_MAX,
2034         .mcgrps         = nbd_mcast_grps,
2035         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2036 };
2037
2038 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2039 {
2040         struct nlattr *dev_opt;
2041         u8 connected = 0;
2042         int ret;
2043
2044         /* This is a little racey, but for status it's ok.  The
2045          * reason we don't take a ref here is because we can't
2046          * take a ref in the index == -1 case as we would need
2047          * to put under the nbd_index_mutex, which could
2048          * deadlock if we are configured to remove ourselves
2049          * once we're disconnected.
2050          */
2051         if (refcount_read(&nbd->config_refs))
2052                 connected = 1;
2053         dev_opt = nla_nest_start(reply, NBD_DEVICE_ITEM);
2054         if (!dev_opt)
2055                 return -EMSGSIZE;
2056         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2057         if (ret)
2058                 return -EMSGSIZE;
2059         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2060                          connected);
2061         if (ret)
2062                 return -EMSGSIZE;
2063         nla_nest_end(reply, dev_opt);
2064         return 0;
2065 }
2066
2067 static int status_cb(int id, void *ptr, void *data)
2068 {
2069         struct nbd_device *nbd = ptr;
2070         return populate_nbd_status(nbd, (struct sk_buff *)data);
2071 }
2072
2073 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2074 {
2075         struct nlattr *dev_list;
2076         struct sk_buff *reply;
2077         void *reply_head;
2078         size_t msg_size;
2079         int index = -1;
2080         int ret = -ENOMEM;
2081
2082         if (info->attrs[NBD_ATTR_INDEX])
2083                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2084
2085         mutex_lock(&nbd_index_mutex);
2086
2087         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2088                                   nla_attr_size(sizeof(u8)));
2089         msg_size *= (index == -1) ? nbd_total_devices : 1;
2090
2091         reply = genlmsg_new(msg_size, GFP_KERNEL);
2092         if (!reply)
2093                 goto out;
2094         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2095                                        NBD_CMD_STATUS);
2096         if (!reply_head) {
2097                 nlmsg_free(reply);
2098                 goto out;
2099         }
2100
2101         dev_list = nla_nest_start(reply, NBD_ATTR_DEVICE_LIST);
2102         if (index == -1) {
2103                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2104                 if (ret) {
2105                         nlmsg_free(reply);
2106                         goto out;
2107                 }
2108         } else {
2109                 struct nbd_device *nbd;
2110                 nbd = idr_find(&nbd_index_idr, index);
2111                 if (nbd) {
2112                         ret = populate_nbd_status(nbd, reply);
2113                         if (ret) {
2114                                 nlmsg_free(reply);
2115                                 goto out;
2116                         }
2117                 }
2118         }
2119         nla_nest_end(reply, dev_list);
2120         genlmsg_end(reply, reply_head);
2121         genlmsg_reply(reply, info);
2122         ret = 0;
2123 out:
2124         mutex_unlock(&nbd_index_mutex);
2125         return ret;
2126 }
2127
2128 static void nbd_connect_reply(struct genl_info *info, int index)
2129 {
2130         struct sk_buff *skb;
2131         void *msg_head;
2132         int ret;
2133
2134         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2135         if (!skb)
2136                 return;
2137         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2138                                      NBD_CMD_CONNECT);
2139         if (!msg_head) {
2140                 nlmsg_free(skb);
2141                 return;
2142         }
2143         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2144         if (ret) {
2145                 nlmsg_free(skb);
2146                 return;
2147         }
2148         genlmsg_end(skb, msg_head);
2149         genlmsg_reply(skb, info);
2150 }
2151
2152 static void nbd_mcast_index(int index)
2153 {
2154         struct sk_buff *skb;
2155         void *msg_head;
2156         int ret;
2157
2158         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2159         if (!skb)
2160                 return;
2161         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2162                                      NBD_CMD_LINK_DEAD);
2163         if (!msg_head) {
2164                 nlmsg_free(skb);
2165                 return;
2166         }
2167         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2168         if (ret) {
2169                 nlmsg_free(skb);
2170                 return;
2171         }
2172         genlmsg_end(skb, msg_head);
2173         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2174 }
2175
2176 static void nbd_dead_link_work(struct work_struct *work)
2177 {
2178         struct link_dead_args *args = container_of(work, struct link_dead_args,
2179                                                    work);
2180         nbd_mcast_index(args->index);
2181         kfree(args);
2182 }
2183
2184 static int __init nbd_init(void)
2185 {
2186         int i;
2187
2188         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2189
2190         if (max_part < 0) {
2191                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2192                 return -EINVAL;
2193         }
2194
2195         part_shift = 0;
2196         if (max_part > 0) {
2197                 part_shift = fls(max_part);
2198
2199                 /*
2200                  * Adjust max_part according to part_shift as it is exported
2201                  * to user space so that user can know the max number of
2202                  * partition kernel should be able to manage.
2203                  *
2204                  * Note that -1 is required because partition 0 is reserved
2205                  * for the whole disk.
2206                  */
2207                 max_part = (1UL << part_shift) - 1;
2208         }
2209
2210         if ((1UL << part_shift) > DISK_MAX_PARTS)
2211                 return -EINVAL;
2212
2213         if (nbds_max > 1UL << (MINORBITS - part_shift))
2214                 return -EINVAL;
2215         recv_workqueue = alloc_workqueue("knbd-recv",
2216                                          WQ_MEM_RECLAIM | WQ_HIGHPRI |
2217                                          WQ_UNBOUND, 0);
2218         if (!recv_workqueue)
2219                 return -ENOMEM;
2220
2221         if (register_blkdev(NBD_MAJOR, "nbd")) {
2222                 destroy_workqueue(recv_workqueue);
2223                 return -EIO;
2224         }
2225
2226         if (genl_register_family(&nbd_genl_family)) {
2227                 unregister_blkdev(NBD_MAJOR, "nbd");
2228                 destroy_workqueue(recv_workqueue);
2229                 return -EINVAL;
2230         }
2231         nbd_dbg_init();
2232
2233         mutex_lock(&nbd_index_mutex);
2234         for (i = 0; i < nbds_max; i++)
2235                 nbd_dev_add(i);
2236         mutex_unlock(&nbd_index_mutex);
2237         return 0;
2238 }
2239
2240 static int nbd_exit_cb(int id, void *ptr, void *data)
2241 {
2242         struct list_head *list = (struct list_head *)data;
2243         struct nbd_device *nbd = ptr;
2244
2245         list_add_tail(&nbd->list, list);
2246         return 0;
2247 }
2248
2249 static void __exit nbd_cleanup(void)
2250 {
2251         struct nbd_device *nbd;
2252         LIST_HEAD(del_list);
2253
2254         nbd_dbg_close();
2255
2256         mutex_lock(&nbd_index_mutex);
2257         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2258         mutex_unlock(&nbd_index_mutex);
2259
2260         while (!list_empty(&del_list)) {
2261                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2262                 list_del_init(&nbd->list);
2263                 if (refcount_read(&nbd->refs) != 1)
2264                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2265                 nbd_put(nbd);
2266         }
2267
2268         idr_destroy(&nbd_index_idr);
2269         genl_unregister_family(&nbd_genl_family);
2270         destroy_workqueue(recv_workqueue);
2271         unregister_blkdev(NBD_MAJOR, "nbd");
2272 }
2273
2274 module_init(nbd_init);
2275 module_exit(nbd_cleanup);
2276
2277 MODULE_DESCRIPTION("Network Block Device");
2278 MODULE_LICENSE("GPL");
2279
2280 module_param(nbds_max, int, 0444);
2281 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2282 module_param(max_part, int, 0444);
2283 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");