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/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/err.h>
  30 #include <linux/kernel.h>
  31 #include <linux/slab.h>
  32 #include <net/sock.h>
  33 #include <linux/net.h>
  34 #include <linux/kthread.h>
  35 #include <linux/types.h>
  36 #include <linux/debugfs.h>
  37 #include <linux/blk-mq.h>
  38
  39 #include <linux/uaccess.h>
  40 #include <asm/types.h>
  41
  42 #include <linux/nbd.h>
  43
  44 struct nbd_sock {
  45         struct socket *sock;
  46         struct mutex tx_lock;
  47 };
  48
  49 #define NBD_TIMEDOUT                    0
  50 #define NBD_DISCONNECT_REQUESTED        1
  51 #define NBD_DISCONNECTED                2
  52 #define NBD_RUNNING                     3
  53
  54 struct nbd_device {
  55         u32 flags;
  56         unsigned long runtime_flags;
  57         struct nbd_sock **socks;
  58         int magic;
  59
  60         struct blk_mq_tag_set tag_set;
  61
  62         struct mutex config_lock;
  63         struct gendisk *disk;
  64         int num_connections;
  65         atomic_t recv_threads;
  66         wait_queue_head_t recv_wq;
  67         loff_t blksize;
  68         loff_t bytesize;
  69
  70         struct task_struct *task_recv;
  71         struct task_struct *task_setup;
  72
  73 #if IS_ENABLED(CONFIG_DEBUG_FS)
  74         struct dentry *dbg_dir;
  75 #endif
  76 };
  77
  78 struct nbd_cmd {
  79         struct nbd_device *nbd;
  80         struct completion send_complete;
  81 };
  82
  83 #if IS_ENABLED(CONFIG_DEBUG_FS)
  84 static struct dentry *nbd_dbg_dir;
  85 #endif
  86
  87 #define nbd_name(nbd) ((nbd)->disk->disk_name)
  88
  89 #define NBD_MAGIC 0x68797548
  90
  91 static unsigned int nbds_max = 16;
  92 static struct nbd_device *nbd_dev;
  93 static int max_part;
  94
  95 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
  96 {
  97         return disk_to_dev(nbd->disk);
  98 }
  99
 100 static bool nbd_is_connected(struct nbd_device *nbd)
 101 {
 102         return !!nbd->task_recv;
 103 }
 104
 105 static const char *nbdcmd_to_ascii(int cmd)
 106 {
 107         switch (cmd) {
 108         case  NBD_CMD_READ: return "read";
 109         case NBD_CMD_WRITE: return "write";
 110         case  NBD_CMD_DISC: return "disconnect";
 111         case NBD_CMD_FLUSH: return "flush";
 112         case  NBD_CMD_TRIM: return "trim/discard";
 113         }
 114         return "invalid";
 115 }
 116
 117 static int nbd_size_clear(struct nbd_device *nbd, struct block_device *bdev)
 118 {
 119         bdev->bd_inode->i_size = 0;
 120         set_capacity(nbd->disk, 0);
 121         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 122
 123         return 0;
 124 }
 125
 126 static void nbd_size_update(struct nbd_device *nbd, struct block_device *bdev)
 127 {
 128         if (!nbd_is_connected(nbd))
 129                 return;
 130
 131         bdev->bd_inode->i_size = nbd->bytesize;
 132         set_capacity(nbd->disk, nbd->bytesize >> 9);
 133         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 134 }
 135
 136 static int nbd_size_set(struct nbd_device *nbd, struct block_device *bdev,
 137                         loff_t blocksize, loff_t nr_blocks)
 138 {
 139         int ret;
 140
 141         ret = set_blocksize(bdev, blocksize);
 142         if (ret)
 143                 return ret;
 144
 145         nbd->blksize = blocksize;
 146         nbd->bytesize = blocksize * nr_blocks;
 147
 148         nbd_size_update(nbd, bdev);
 149
 150         return 0;
 151 }
 152
 153 static void nbd_end_request(struct nbd_cmd *cmd)
 154 {
 155         struct nbd_device *nbd = cmd->nbd;
 156         struct request *req = blk_mq_rq_from_pdu(cmd);
 157         int error = req->errors ? -EIO : 0;
 158
 159         dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", cmd,
 160                 error ? "failed" : "done");
 161
 162         blk_mq_complete_request(req, error);
 163 }
 164
 165 /*
 166  * Forcibly shutdown the socket causing all listeners to error
 167  */
 168 static void sock_shutdown(struct nbd_device *nbd)
 169 {
 170         int i;
 171
 172         if (nbd->num_connections == 0)
 173                 return;
 174         if (test_and_set_bit(NBD_DISCONNECTED, &nbd->runtime_flags))
 175                 return;
 176
 177         for (i = 0; i < nbd->num_connections; i++) {
 178                 struct nbd_sock *nsock = nbd->socks[i];
 179                 mutex_lock(&nsock->tx_lock);
 180                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 181                 mutex_unlock(&nsock->tx_lock);
 182         }
 183         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 184 }
 185
 186 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 187                                                  bool reserved)
 188 {
 189         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 190         struct nbd_device *nbd = cmd->nbd;
 191
 192         dev_err(nbd_to_dev(nbd), "Connection timed out, shutting down connection\n");
 193         set_bit(NBD_TIMEDOUT, &nbd->runtime_flags);
 194         req->errors++;
 195
 196         /*
 197          * If our disconnect packet times out then we're already holding the
 198          * config_lock and could deadlock here, so just set an error and return,
 199          * we'll handle shutting everything down later.
 200          */
 201         if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 202                 return BLK_EH_HANDLED;
 203         mutex_lock(&nbd->config_lock);
 204         sock_shutdown(nbd);
 205         mutex_unlock(&nbd->config_lock);
 206         return BLK_EH_HANDLED;
 207 }
 208
 209 /*
 210  *  Send or receive packet.
 211  */
 212 static int sock_xmit(struct nbd_device *nbd, int index, int send, void *buf,
 213                      int size, int msg_flags)
 214 {
 215         struct socket *sock = nbd->socks[index]->sock;
 216         int result;
 217         struct msghdr msg;
 218         struct kvec iov;
 219         unsigned long pflags = current->flags;
 220
 221         if (unlikely(!sock)) {
 222                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 223                         "Attempted %s on closed socket in sock_xmit\n",
 224                         (send ? "send" : "recv"));
 225                 return -EINVAL;
 226         }
 227
 228         current->flags |= PF_MEMALLOC;
 229         do {
 230                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 231                 iov.iov_base = buf;
 232                 iov.iov_len = size;
 233                 msg.msg_name = NULL;
 234                 msg.msg_namelen = 0;
 235                 msg.msg_control = NULL;
 236                 msg.msg_controllen = 0;
 237                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 238
 239                 if (send)
 240                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 241                 else
 242                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 243                                                 msg.msg_flags);
 244
 245                 if (result <= 0) {
 246                         if (result == 0)
 247                                 result = -EPIPE; /* short read */
 248                         break;
 249                 }
 250                 size -= result;
 251                 buf += result;
 252         } while (size > 0);
 253
 254         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 255
 256         return result;
 257 }
 258
 259 static inline int sock_send_bvec(struct nbd_device *nbd, int index,
 260                                  struct bio_vec *bvec, int flags)
 261 {
 262         int result;
 263         void *kaddr = kmap(bvec->bv_page);
 264         result = sock_xmit(nbd, index, 1, kaddr + bvec->bv_offset,
 265                            bvec->bv_len, flags);
 266         kunmap(bvec->bv_page);
 267         return result;
 268 }
 269
 270 /* always call with the tx_lock held */
 271 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 272 {
 273         struct request *req = blk_mq_rq_from_pdu(cmd);
 274         int result, flags;
 275         struct nbd_request request;
 276         unsigned long size = blk_rq_bytes(req);
 277         struct bio *bio;
 278         u32 type;
 279         u32 tag = blk_mq_unique_tag(req);
 280
 281         if (req_op(req) == REQ_OP_DISCARD)
 282                 type = NBD_CMD_TRIM;
 283         else if (req_op(req) == REQ_OP_FLUSH)
 284                 type = NBD_CMD_FLUSH;
 285         else if (rq_data_dir(req) == WRITE)
 286                 type = NBD_CMD_WRITE;
 287         else
 288                 type = NBD_CMD_READ;
 289
 290         memset(&request, 0, sizeof(request));
 291         request.magic = htonl(NBD_REQUEST_MAGIC);
 292         request.type = htonl(type);
 293         if (type != NBD_CMD_FLUSH) {
 294                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 295                 request.len = htonl(size);
 296         }
 297         memcpy(request.handle, &tag, sizeof(tag));
 298
 299         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 300                 cmd, nbdcmd_to_ascii(type),
 301                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 302         result = sock_xmit(nbd, index, 1, &request, sizeof(request),
 303                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
 304         if (result <= 0) {
 305                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 306                         "Send control failed (result %d)\n", result);
 307                 return -EIO;
 308         }
 309
 310         if (type != NBD_CMD_WRITE)
 311                 return 0;
 312
 313         flags = 0;
 314         bio = req->bio;
 315         while (bio) {
 316                 struct bio *next = bio->bi_next;
 317                 struct bvec_iter iter;
 318                 struct bio_vec bvec;
 319
 320                 bio_for_each_segment(bvec, bio, iter) {
 321                         bool is_last = !next && bio_iter_last(bvec, iter);
 322
 323                         if (is_last)
 324                                 flags = MSG_MORE;
 325                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 326                                 cmd, bvec.bv_len);
 327                         result = sock_send_bvec(nbd, index, &bvec, flags);
 328                         if (result <= 0) {
 329                                 dev_err(disk_to_dev(nbd->disk),
 330                                         "Send data failed (result %d)\n",
 331                                         result);
 332                                 return -EIO;
 333                         }
 334                         /*
 335                          * The completion might already have come in,
 336                          * so break for the last one instead of letting
 337                          * the iterator do it. This prevents use-after-free
 338                          * of the bio.
 339                          */
 340                         if (is_last)
 341                                 break;
 342                 }
 343                 bio = next;
 344         }
 345         return 0;
 346 }
 347
 348 static inline int sock_recv_bvec(struct nbd_device *nbd, int index,
 349                                  struct bio_vec *bvec)
 350 {
 351         int result;
 352         void *kaddr = kmap(bvec->bv_page);
 353         result = sock_xmit(nbd, index, 0, kaddr + bvec->bv_offset,
 354                            bvec->bv_len, MSG_WAITALL);
 355         kunmap(bvec->bv_page);
 356         return result;
 357 }
 358
 359 /* NULL returned = something went wrong, inform userspace */
 360 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 361 {
 362         int result;
 363         struct nbd_reply reply;
 364         struct nbd_cmd *cmd;
 365         struct request *req = NULL;
 366         u16 hwq;
 367         u32 tag;
 368
 369         reply.magic = 0;
 370         result = sock_xmit(nbd, index, 0, &reply, sizeof(reply), MSG_WAITALL);
 371         if (result <= 0) {
 372                 if (!test_bit(NBD_DISCONNECTED, &nbd->runtime_flags) &&
 373                     !test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags))
 374                         dev_err(disk_to_dev(nbd->disk),
 375                                 "Receive control failed (result %d)\n", result);
 376                 return ERR_PTR(result);
 377         }
 378
 379         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 380                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 381                                 (unsigned long)ntohl(reply.magic));
 382                 return ERR_PTR(-EPROTO);
 383         }
 384
 385         memcpy(&tag, reply.handle, sizeof(u32));
 386
 387         hwq = blk_mq_unique_tag_to_hwq(tag);
 388         if (hwq < nbd->tag_set.nr_hw_queues)
 389                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 390                                        blk_mq_unique_tag_to_tag(tag));
 391         if (!req || !blk_mq_request_started(req)) {
 392                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 393                         tag, req);
 394                 return ERR_PTR(-ENOENT);
 395         }
 396         cmd = blk_mq_rq_to_pdu(req);
 397         if (ntohl(reply.error)) {
 398                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 399                         ntohl(reply.error));
 400                 req->errors++;
 401                 return cmd;
 402         }
 403
 404         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", cmd);
 405         if (rq_data_dir(req) != WRITE) {
 406                 struct req_iterator iter;
 407                 struct bio_vec bvec;
 408
 409                 rq_for_each_segment(bvec, req, iter) {
 410                         result = sock_recv_bvec(nbd, index, &bvec);
 411                         if (result <= 0) {
 412                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 413                                         result);
 414                                 req->errors++;
 415                                 return cmd;
 416                         }
 417                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 418                                 cmd, bvec.bv_len);
 419                 }
 420         } else {
 421                 /* See the comment in nbd_queue_rq. */
 422                 wait_for_completion(&cmd->send_complete);
 423         }
 424         return cmd;
 425 }
 426
 427 static ssize_t pid_show(struct device *dev,
 428                         struct device_attribute *attr, char *buf)
 429 {
 430         struct gendisk *disk = dev_to_disk(dev);
 431         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 432
 433         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 434 }
 435
 436 static struct device_attribute pid_attr = {
 437         .attr = { .name = "pid", .mode = S_IRUGO},
 438         .show = pid_show,
 439 };
 440
 441 struct recv_thread_args {
 442         struct work_struct work;
 443         struct nbd_device *nbd;
 444         int index;
 445 };
 446
 447 static void recv_work(struct work_struct *work)
 448 {
 449         struct recv_thread_args *args = container_of(work,
 450                                                      struct recv_thread_args,
 451                                                      work);
 452         struct nbd_device *nbd = args->nbd;
 453         struct nbd_cmd *cmd;
 454         int ret = 0;
 455
 456         BUG_ON(nbd->magic != NBD_MAGIC);
 457         while (1) {
 458                 cmd = nbd_read_stat(nbd, args->index);
 459                 if (IS_ERR(cmd)) {
 460                         ret = PTR_ERR(cmd);
 461                         break;
 462                 }
 463
 464                 nbd_end_request(cmd);
 465         }
 466
 467         /*
 468          * We got an error, shut everybody down if this wasn't the result of a
 469          * disconnect request.
 470          */
 471         if (ret && !test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags))
 472                 sock_shutdown(nbd);
 473         atomic_dec(&nbd->recv_threads);
 474         wake_up(&nbd->recv_wq);
 475 }
 476
 477 static void nbd_clear_req(struct request *req, void *data, bool reserved)
 478 {
 479         struct nbd_cmd *cmd;
 480
 481         if (!blk_mq_request_started(req))
 482                 return;
 483         cmd = blk_mq_rq_to_pdu(req);
 484         req->errors++;
 485         nbd_end_request(cmd);
 486 }
 487
 488 static void nbd_clear_que(struct nbd_device *nbd)
 489 {
 490         BUG_ON(nbd->magic != NBD_MAGIC);
 491
 492         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 493         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 494 }
 495
 496
 497 static void nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 498 {
 499         struct request *req = blk_mq_rq_from_pdu(cmd);
 500         struct nbd_device *nbd = cmd->nbd;
 501         struct nbd_sock *nsock;
 502
 503         if (index >= nbd->num_connections) {
 504                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 505                                     "Attempted send on invalid socket\n");
 506                 goto error_out;
 507         }
 508
 509         if (test_bit(NBD_DISCONNECTED, &nbd->runtime_flags)) {
 510                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 511                                     "Attempted send on closed socket\n");
 512                 goto error_out;
 513         }
 514
 515         if (req->cmd_type != REQ_TYPE_FS &&
 516             req->cmd_type != REQ_TYPE_DRV_PRIV)
 517                 goto error_out;
 518
 519         if (req->cmd_type == REQ_TYPE_FS &&
 520             rq_data_dir(req) == WRITE &&
 521             (nbd->flags & NBD_FLAG_READ_ONLY)) {
 522                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 523                                     "Write on read-only\n");
 524                 goto error_out;
 525         }
 526
 527         req->errors = 0;
 528
 529         nsock = nbd->socks[index];
 530         mutex_lock(&nsock->tx_lock);
 531         if (unlikely(!nsock->sock)) {
 532                 mutex_unlock(&nsock->tx_lock);
 533                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 534                                     "Attempted send on closed socket\n");
 535                 goto error_out;
 536         }
 537
 538         if (nbd_send_cmd(nbd, cmd, index) != 0) {
 539                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 540                                     "Request send failed\n");
 541                 req->errors++;
 542                 nbd_end_request(cmd);
 543         }
 544
 545         mutex_unlock(&nsock->tx_lock);
 546
 547         return;
 548
 549 error_out:
 550         req->errors++;
 551         nbd_end_request(cmd);
 552 }
 553
 554 static int nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 555                         const struct blk_mq_queue_data *bd)
 556 {
 557         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 558
 559         /*
 560          * Since we look at the bio's to send the request over the network we
 561          * need to make sure the completion work doesn't mark this request done
 562          * before we are done doing our send.  This keeps us from dereferencing
 563          * freed data if we have particularly fast completions (ie we get the
 564          * completion before we exit sock_xmit on the last bvec) or in the case
 565          * that the server is misbehaving (or there was an error) before we're
 566          * done sending everything over the wire.
 567          */
 568         init_completion(&cmd->send_complete);
 569         blk_mq_start_request(bd->rq);
 570         nbd_handle_cmd(cmd, hctx->queue_num);
 571         complete(&cmd->send_complete);
 572
 573         return BLK_MQ_RQ_QUEUE_OK;
 574 }
 575
 576 static int nbd_add_socket(struct nbd_device *nbd, struct socket *sock)
 577 {
 578         struct nbd_sock **socks;
 579         struct nbd_sock *nsock;
 580
 581         if (!nbd->task_setup)
 582                 nbd->task_setup = current;
 583         if (nbd->task_setup != current) {
 584                 dev_err(disk_to_dev(nbd->disk),
 585                         "Device being setup by another task");
 586                 return -EINVAL;
 587         }
 588
 589         socks = krealloc(nbd->socks, (nbd->num_connections + 1) *
 590                          sizeof(struct nbd_sock *), GFP_KERNEL);
 591         if (!socks)
 592                 return -ENOMEM;
 593         nsock = kzalloc(sizeof(struct nbd_sock), GFP_KERNEL);
 594         if (!nsock)
 595                 return -ENOMEM;
 596
 597         nbd->socks = socks;
 598
 599         mutex_init(&nsock->tx_lock);
 600         nsock->sock = sock;
 601         socks[nbd->num_connections++] = nsock;
 602
 603         return 0;
 604 }
 605
 606 /* Reset all properties of an NBD device */
 607 static void nbd_reset(struct nbd_device *nbd)
 608 {
 609         int i;
 610
 611         for (i = 0; i < nbd->num_connections; i++)
 612                 kfree(nbd->socks[i]);
 613         kfree(nbd->socks);
 614         nbd->socks = NULL;
 615         nbd->runtime_flags = 0;
 616         nbd->blksize = 1024;
 617         nbd->bytesize = 0;
 618         set_capacity(nbd->disk, 0);
 619         nbd->flags = 0;
 620         nbd->tag_set.timeout = 0;
 621         nbd->num_connections = 0;
 622         nbd->task_setup = NULL;
 623         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 624 }
 625
 626 static void nbd_bdev_reset(struct block_device *bdev)
 627 {
 628         set_device_ro(bdev, false);
 629         bdev->bd_inode->i_size = 0;
 630         if (max_part > 0) {
 631                 blkdev_reread_part(bdev);
 632                 bdev->bd_invalidated = 1;
 633         }
 634 }
 635
 636 static void nbd_parse_flags(struct nbd_device *nbd, struct block_device *bdev)
 637 {
 638         if (nbd->flags & NBD_FLAG_READ_ONLY)
 639                 set_device_ro(bdev, true);
 640         if (nbd->flags & NBD_FLAG_SEND_TRIM)
 641                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 642         if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 643                 blk_queue_write_cache(nbd->disk->queue, true, false);
 644         else
 645                 blk_queue_write_cache(nbd->disk->queue, false, false);
 646 }
 647
 648 static void send_disconnects(struct nbd_device *nbd)
 649 {
 650         struct nbd_request request = {};
 651         int i, ret;
 652
 653         request.magic = htonl(NBD_REQUEST_MAGIC);
 654         request.type = htonl(NBD_CMD_DISC);
 655
 656         for (i = 0; i < nbd->num_connections; i++) {
 657                 ret = sock_xmit(nbd, i, 1, &request, sizeof(request), 0);
 658                 if (ret <= 0)
 659                         dev_err(disk_to_dev(nbd->disk),
 660                                 "Send disconnect failed %d\n", ret);
 661         }
 662 }
 663
 664 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 665 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 666
 667 /* Must be called with config_lock held */
 668 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 669                        unsigned int cmd, unsigned long arg)
 670 {
 671         switch (cmd) {
 672         case NBD_DISCONNECT: {
 673                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 674                 if (!nbd->socks)
 675                         return -EINVAL;
 676
 677                 mutex_unlock(&nbd->config_lock);
 678                 fsync_bdev(bdev);
 679                 mutex_lock(&nbd->config_lock);
 680
 681                 /* Check again after getting mutex back.  */
 682                 if (!nbd->socks)
 683                         return -EINVAL;
 684
 685                 if (!test_and_set_bit(NBD_DISCONNECT_REQUESTED,
 686                                       &nbd->runtime_flags))
 687                         send_disconnects(nbd);
 688                 return 0;
 689         }
 690
 691         case NBD_CLEAR_SOCK:
 692                 sock_shutdown(nbd);
 693                 nbd_clear_que(nbd);
 694                 kill_bdev(bdev);
 695                 nbd_bdev_reset(bdev);
 696                 /*
 697                  * We want to give the run thread a chance to wait for everybody
 698                  * to clean up and then do it's own cleanup.
 699                  */
 700                 if (!test_bit(NBD_RUNNING, &nbd->runtime_flags)) {
 701                         int i;
 702
 703                         for (i = 0; i < nbd->num_connections; i++)
 704                                 kfree(nbd->socks[i]);
 705                         kfree(nbd->socks);
 706                         nbd->socks = NULL;
 707                         nbd->num_connections = 0;
 708                         nbd->task_setup = NULL;
 709                 }
 710                 return 0;
 711
 712         case NBD_SET_SOCK: {
 713                 int err;
 714                 struct socket *sock = sockfd_lookup(arg, &err);
 715
 716                 if (!sock)
 717                         return err;
 718
 719                 err = nbd_add_socket(nbd, sock);
 720                 if (!err && max_part)
 721                         bdev->bd_invalidated = 1;
 722
 723                 return err;
 724         }
 725
 726         case NBD_SET_BLKSIZE: {
 727                 loff_t bsize = div_s64(nbd->bytesize, arg);
 728
 729                 return nbd_size_set(nbd, bdev, arg, bsize);
 730         }
 731
 732         case NBD_SET_SIZE:
 733                 return nbd_size_set(nbd, bdev, nbd->blksize,
 734                                         div_s64(arg, nbd->blksize));
 735
 736         case NBD_SET_SIZE_BLOCKS:
 737                 return nbd_size_set(nbd, bdev, nbd->blksize, arg);
 738
 739         case NBD_SET_TIMEOUT:
 740                 nbd->tag_set.timeout = arg * HZ;
 741                 return 0;
 742
 743         case NBD_SET_FLAGS:
 744                 nbd->flags = arg;
 745                 return 0;
 746
 747         case NBD_DO_IT: {
 748                 struct recv_thread_args *args;
 749                 int num_connections = nbd->num_connections;
 750                 int error = 0, i;
 751
 752                 if (nbd->task_recv)
 753                         return -EBUSY;
 754                 if (!nbd->socks)
 755                         return -EINVAL;
 756                 if (num_connections > 1 &&
 757                     !(nbd->flags & NBD_FLAG_CAN_MULTI_CONN)) {
 758                         dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
 759                         error = -EINVAL;
 760                         goto out_err;
 761                 }
 762
 763                 set_bit(NBD_RUNNING, &nbd->runtime_flags);
 764                 blk_mq_update_nr_hw_queues(&nbd->tag_set, nbd->num_connections);
 765                 args = kcalloc(num_connections, sizeof(*args), GFP_KERNEL);
 766                 if (!args) {
 767                         error = -ENOMEM;
 768                         goto out_err;
 769                 }
 770                 nbd->task_recv = current;
 771                 mutex_unlock(&nbd->config_lock);
 772
 773                 nbd_parse_flags(nbd, bdev);
 774
 775                 error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 776                 if (error) {
 777                         dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 778                         goto out_recv;
 779                 }
 780
 781                 nbd_size_update(nbd, bdev);
 782
 783                 nbd_dev_dbg_init(nbd);
 784                 for (i = 0; i < num_connections; i++) {
 785                         sk_set_memalloc(nbd->socks[i]->sock->sk);
 786                         atomic_inc(&nbd->recv_threads);
 787                         INIT_WORK(&args[i].work, recv_work);
 788                         args[i].nbd = nbd;
 789                         args[i].index = i;
 790                         queue_work(system_long_wq, &args[i].work);
 791                 }
 792                 wait_event_interruptible(nbd->recv_wq,
 793                                          atomic_read(&nbd->recv_threads) == 0);
 794                 for (i = 0; i < num_connections; i++)
 795                         flush_work(&args[i].work);
 796                 nbd_dev_dbg_close(nbd);
 797                 nbd_size_clear(nbd, bdev);
 798                 device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 799 out_recv:
 800                 mutex_lock(&nbd->config_lock);
 801                 nbd->task_recv = NULL;
 802 out_err:
 803                 sock_shutdown(nbd);
 804                 nbd_clear_que(nbd);
 805                 kill_bdev(bdev);
 806                 nbd_bdev_reset(bdev);
 807
 808                 /* user requested, ignore socket errors */
 809                 if (test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags))
 810                         error = 0;
 811                 if (test_bit(NBD_TIMEDOUT, &nbd->runtime_flags))
 812                         error = -ETIMEDOUT;
 813
 814                 nbd_reset(nbd);
 815                 return error;
 816         }
 817
 818         case NBD_CLEAR_QUE:
 819                 /*
 820                  * This is for compatibility only.  The queue is always cleared
 821                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 822                  */
 823                 return 0;
 824
 825         case NBD_PRINT_DEBUG:
 826                 /*
 827                  * For compatibility only, we no longer keep a list of
 828                  * outstanding requests.
 829                  */
 830                 return 0;
 831         }
 832         return -ENOTTY;
 833 }
 834
 835 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 836                      unsigned int cmd, unsigned long arg)
 837 {
 838         struct nbd_device *nbd = bdev->bd_disk->private_data;
 839         int error;
 840
 841         if (!capable(CAP_SYS_ADMIN))
 842                 return -EPERM;
 843
 844         BUG_ON(nbd->magic != NBD_MAGIC);
 845
 846         mutex_lock(&nbd->config_lock);
 847         error = __nbd_ioctl(bdev, nbd, cmd, arg);
 848         mutex_unlock(&nbd->config_lock);
 849
 850         return error;
 851 }
 852
 853 static const struct block_device_operations nbd_fops =
 854 {
 855         .owner =        THIS_MODULE,
 856         .ioctl =        nbd_ioctl,
 857         .compat_ioctl = nbd_ioctl,
 858 };
 859
 860 #if IS_ENABLED(CONFIG_DEBUG_FS)
 861
 862 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
 863 {
 864         struct nbd_device *nbd = s->private;
 865
 866         if (nbd->task_recv)
 867                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
 868
 869         return 0;
 870 }
 871
 872 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
 873 {
 874         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
 875 }
 876
 877 static const struct file_operations nbd_dbg_tasks_ops = {
 878         .open = nbd_dbg_tasks_open,
 879         .read = seq_read,
 880         .llseek = seq_lseek,
 881         .release = single_release,
 882 };
 883
 884 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
 885 {
 886         struct nbd_device *nbd = s->private;
 887         u32 flags = nbd->flags;
 888
 889         seq_printf(s, "Hex: 0x%08x\n\n", flags);
 890
 891         seq_puts(s, "Known flags:\n");
 892
 893         if (flags & NBD_FLAG_HAS_FLAGS)
 894                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
 895         if (flags & NBD_FLAG_READ_ONLY)
 896                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
 897         if (flags & NBD_FLAG_SEND_FLUSH)
 898                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
 899         if (flags & NBD_FLAG_SEND_TRIM)
 900                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
 901
 902         return 0;
 903 }
 904
 905 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
 906 {
 907         return single_open(file, nbd_dbg_flags_show, inode->i_private);
 908 }
 909
 910 static const struct file_operations nbd_dbg_flags_ops = {
 911         .open = nbd_dbg_flags_open,
 912         .read = seq_read,
 913         .llseek = seq_lseek,
 914         .release = single_release,
 915 };
 916
 917 static int nbd_dev_dbg_init(struct nbd_device *nbd)
 918 {
 919         struct dentry *dir;
 920
 921         if (!nbd_dbg_dir)
 922                 return -EIO;
 923
 924         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
 925         if (!dir) {
 926                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
 927                         nbd_name(nbd));
 928                 return -EIO;
 929         }
 930         nbd->dbg_dir = dir;
 931
 932         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
 933         debugfs_create_u64("size_bytes", 0444, dir, &nbd->bytesize);
 934         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
 935         debugfs_create_u64("blocksize", 0444, dir, &nbd->blksize);
 936         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
 937
 938         return 0;
 939 }
 940
 941 static void nbd_dev_dbg_close(struct nbd_device *nbd)
 942 {
 943         debugfs_remove_recursive(nbd->dbg_dir);
 944 }
 945
 946 static int nbd_dbg_init(void)
 947 {
 948         struct dentry *dbg_dir;
 949
 950         dbg_dir = debugfs_create_dir("nbd", NULL);
 951         if (!dbg_dir)
 952                 return -EIO;
 953
 954         nbd_dbg_dir = dbg_dir;
 955
 956         return 0;
 957 }
 958
 959 static void nbd_dbg_close(void)
 960 {
 961         debugfs_remove_recursive(nbd_dbg_dir);
 962 }
 963
 964 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
 965
 966 static int nbd_dev_dbg_init(struct nbd_device *nbd)
 967 {
 968         return 0;
 969 }
 970
 971 static void nbd_dev_dbg_close(struct nbd_device *nbd)
 972 {
 973 }
 974
 975 static int nbd_dbg_init(void)
 976 {
 977         return 0;
 978 }
 979
 980 static void nbd_dbg_close(void)
 981 {
 982 }
 983
 984 #endif
 985
 986 static int nbd_init_request(void *data, struct request *rq,
 987                             unsigned int hctx_idx, unsigned int request_idx,
 988                             unsigned int numa_node)
 989 {
 990         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
 991         cmd->nbd = data;
 992         return 0;
 993 }
 994
 995 static struct blk_mq_ops nbd_mq_ops = {
 996         .queue_rq       = nbd_queue_rq,
 997         .init_request   = nbd_init_request,
 998         .timeout        = nbd_xmit_timeout,
 999 };
1000
1001 /*
1002  * And here should be modules and kernel interface
1003  *  (Just smiley confuses emacs :-)
1004  */
1005
1006 static int __init nbd_init(void)
1007 {
1008         int err = -ENOMEM;
1009         int i;
1010         int part_shift;
1011
1012         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
1013
1014         if (max_part < 0) {
1015                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
1016                 return -EINVAL;
1017         }
1018
1019         part_shift = 0;
1020         if (max_part > 0) {
1021                 part_shift = fls(max_part);
1022
1023                 /*
1024                  * Adjust max_part according to part_shift as it is exported
1025                  * to user space so that user can know the max number of
1026                  * partition kernel should be able to manage.
1027                  *
1028                  * Note that -1 is required because partition 0 is reserved
1029                  * for the whole disk.
1030                  */
1031                 max_part = (1UL << part_shift) - 1;
1032         }
1033
1034         if ((1UL << part_shift) > DISK_MAX_PARTS)
1035                 return -EINVAL;
1036
1037         if (nbds_max > 1UL << (MINORBITS - part_shift))
1038                 return -EINVAL;
1039
1040         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
1041         if (!nbd_dev)
1042                 return -ENOMEM;
1043
1044         for (i = 0; i < nbds_max; i++) {
1045                 struct gendisk *disk = alloc_disk(1 << part_shift);
1046                 if (!disk)
1047                         goto out;
1048                 nbd_dev[i].disk = disk;
1049
1050                 nbd_dev[i].tag_set.ops = &nbd_mq_ops;
1051                 nbd_dev[i].tag_set.nr_hw_queues = 1;
1052                 nbd_dev[i].tag_set.queue_depth = 128;
1053                 nbd_dev[i].tag_set.numa_node = NUMA_NO_NODE;
1054                 nbd_dev[i].tag_set.cmd_size = sizeof(struct nbd_cmd);
1055                 nbd_dev[i].tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1056                         BLK_MQ_F_SG_MERGE | BLK_MQ_F_BLOCKING;
1057                 nbd_dev[i].tag_set.driver_data = &nbd_dev[i];
1058
1059                 err = blk_mq_alloc_tag_set(&nbd_dev[i].tag_set);
1060                 if (err) {
1061                         put_disk(disk);
1062                         goto out;
1063                 }
1064
1065                 /*
1066                  * The new linux 2.5 block layer implementation requires
1067                  * every gendisk to have its very own request_queue struct.
1068                  * These structs are big so we dynamically allocate them.
1069                  */
1070                 disk->queue = blk_mq_init_queue(&nbd_dev[i].tag_set);
1071                 if (!disk->queue) {
1072                         blk_mq_free_tag_set(&nbd_dev[i].tag_set);
1073                         put_disk(disk);
1074                         goto out;
1075                 }
1076
1077                 /*
1078                  * Tell the block layer that we are not a rotational device
1079                  */
1080                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
1081                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1082                 disk->queue->limits.discard_granularity = 512;
1083                 blk_queue_max_discard_sectors(disk->queue, UINT_MAX);
1084                 disk->queue->limits.discard_zeroes_data = 0;
1085                 blk_queue_max_hw_sectors(disk->queue, 65536);
1086                 disk->queue->limits.max_sectors = 256;
1087         }
1088
1089         if (register_blkdev(NBD_MAJOR, "nbd")) {
1090                 err = -EIO;
1091                 goto out;
1092         }
1093
1094         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
1095
1096         nbd_dbg_init();
1097
1098         for (i = 0; i < nbds_max; i++) {
1099                 struct gendisk *disk = nbd_dev[i].disk;
1100                 nbd_dev[i].magic = NBD_MAGIC;
1101                 mutex_init(&nbd_dev[i].config_lock);
1102                 disk->major = NBD_MAJOR;
1103                 disk->first_minor = i << part_shift;
1104                 disk->fops = &nbd_fops;
1105                 disk->private_data = &nbd_dev[i];
1106                 sprintf(disk->disk_name, "nbd%d", i);
1107                 init_waitqueue_head(&nbd_dev[i].recv_wq);
1108                 nbd_reset(&nbd_dev[i]);
1109                 add_disk(disk);
1110         }
1111
1112         return 0;
1113 out:
1114         while (i--) {
1115                 blk_mq_free_tag_set(&nbd_dev[i].tag_set);
1116                 blk_cleanup_queue(nbd_dev[i].disk->queue);
1117                 put_disk(nbd_dev[i].disk);
1118         }
1119         kfree(nbd_dev);
1120         return err;
1121 }
1122
1123 static void __exit nbd_cleanup(void)
1124 {
1125         int i;
1126
1127         nbd_dbg_close();
1128
1129         for (i = 0; i < nbds_max; i++) {
1130                 struct gendisk *disk = nbd_dev[i].disk;
1131                 nbd_dev[i].magic = 0;
1132                 if (disk) {
1133                         del_gendisk(disk);
1134                         blk_cleanup_queue(disk->queue);
1135                         blk_mq_free_tag_set(&nbd_dev[i].tag_set);
1136                         put_disk(disk);
1137                 }
1138         }
1139         unregister_blkdev(NBD_MAJOR, "nbd");
1140         kfree(nbd_dev);
1141         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
1142 }
1143
1144 module_init(nbd_init);
1145 module_exit(nbd_cleanup);
1146
1147 MODULE_DESCRIPTION("Network Block Device");
1148 MODULE_LICENSE("GPL");
1149
1150 module_param(nbds_max, int, 0444);
1151 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
1152 module_param(max_part, int, 0444);
1153 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");