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 static struct workqueue_struct *recv_workqueue;
  95
  96 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
  97 {
  98         return disk_to_dev(nbd->disk);
  99 }
 100
 101 static bool nbd_is_connected(struct nbd_device *nbd)
 102 {
 103         return !!nbd->task_recv;
 104 }
 105
 106 static const char *nbdcmd_to_ascii(int cmd)
 107 {
 108         switch (cmd) {
 109         case  NBD_CMD_READ: return "read";
 110         case NBD_CMD_WRITE: return "write";
 111         case  NBD_CMD_DISC: return "disconnect";
 112         case NBD_CMD_FLUSH: return "flush";
 113         case  NBD_CMD_TRIM: return "trim/discard";
 114         }
 115         return "invalid";
 116 }
 117
 118 static int nbd_size_clear(struct nbd_device *nbd, struct block_device *bdev)
 119 {
 120         bdev->bd_inode->i_size = 0;
 121         set_capacity(nbd->disk, 0);
 122         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 123
 124         return 0;
 125 }
 126
 127 static void nbd_size_update(struct nbd_device *nbd, struct block_device *bdev)
 128 {
 129         if (!nbd_is_connected(nbd))
 130                 return;
 131
 132         bdev->bd_inode->i_size = nbd->bytesize;
 133         set_capacity(nbd->disk, nbd->bytesize >> 9);
 134         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 135 }
 136
 137 static int nbd_size_set(struct nbd_device *nbd, struct block_device *bdev,
 138                         loff_t blocksize, loff_t nr_blocks)
 139 {
 140         int ret;
 141
 142         ret = set_blocksize(bdev, blocksize);
 143         if (ret)
 144                 return ret;
 145
 146         nbd->blksize = blocksize;
 147         nbd->bytesize = blocksize * nr_blocks;
 148
 149         nbd_size_update(nbd, bdev);
 150
 151         return 0;
 152 }
 153
 154 static void nbd_end_request(struct nbd_cmd *cmd)
 155 {
 156         struct nbd_device *nbd = cmd->nbd;
 157         struct request *req = blk_mq_rq_from_pdu(cmd);
 158         int error = req->errors ? -EIO : 0;
 159
 160         dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", cmd,
 161                 error ? "failed" : "done");
 162
 163         blk_mq_complete_request(req, error);
 164 }
 165
 166 /*
 167  * Forcibly shutdown the socket causing all listeners to error
 168  */
 169 static void sock_shutdown(struct nbd_device *nbd)
 170 {
 171         int i;
 172
 173         if (nbd->num_connections == 0)
 174                 return;
 175         if (test_and_set_bit(NBD_DISCONNECTED, &nbd->runtime_flags))
 176                 return;
 177
 178         for (i = 0; i < nbd->num_connections; i++) {
 179                 struct nbd_sock *nsock = nbd->socks[i];
 180                 mutex_lock(&nsock->tx_lock);
 181                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 182                 mutex_unlock(&nsock->tx_lock);
 183         }
 184         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 185 }
 186
 187 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 188                                                  bool reserved)
 189 {
 190         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 191         struct nbd_device *nbd = cmd->nbd;
 192
 193         dev_err(nbd_to_dev(nbd), "Connection timed out, shutting down connection\n");
 194         set_bit(NBD_TIMEDOUT, &nbd->runtime_flags);
 195         req->errors++;
 196
 197         mutex_lock(&nbd->config_lock);
 198         sock_shutdown(nbd);
 199         mutex_unlock(&nbd->config_lock);
 200         return BLK_EH_HANDLED;
 201 }
 202
 203 /*
 204  *  Send or receive packet.
 205  */
 206 static int sock_xmit(struct nbd_device *nbd, int index, int send, void *buf,
 207                      int size, int msg_flags)
 208 {
 209         struct socket *sock = nbd->socks[index]->sock;
 210         int result;
 211         struct msghdr msg;
 212         struct kvec iov;
 213         unsigned long pflags = current->flags;
 214
 215         if (unlikely(!sock)) {
 216                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 217                         "Attempted %s on closed socket in sock_xmit\n",
 218                         (send ? "send" : "recv"));
 219                 return -EINVAL;
 220         }
 221
 222         current->flags |= PF_MEMALLOC;
 223         do {
 224                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 225                 iov.iov_base = buf;
 226                 iov.iov_len = size;
 227                 msg.msg_name = NULL;
 228                 msg.msg_namelen = 0;
 229                 msg.msg_control = NULL;
 230                 msg.msg_controllen = 0;
 231                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 232
 233                 if (send)
 234                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 235                 else
 236                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 237                                                 msg.msg_flags);
 238
 239                 if (result <= 0) {
 240                         if (result == 0)
 241                                 result = -EPIPE; /* short read */
 242                         break;
 243                 }
 244                 size -= result;
 245                 buf += result;
 246         } while (size > 0);
 247
 248         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 249
 250         return result;
 251 }
 252
 253 static inline int sock_send_bvec(struct nbd_device *nbd, int index,
 254                                  struct bio_vec *bvec, int flags)
 255 {
 256         int result;
 257         void *kaddr = kmap(bvec->bv_page);
 258         result = sock_xmit(nbd, index, 1, kaddr + bvec->bv_offset,
 259                            bvec->bv_len, flags);
 260         kunmap(bvec->bv_page);
 261         return result;
 262 }
 263
 264 /* always call with the tx_lock held */
 265 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 266 {
 267         struct request *req = blk_mq_rq_from_pdu(cmd);
 268         int result;
 269         struct nbd_request request;
 270         unsigned long size = blk_rq_bytes(req);
 271         struct bio *bio;
 272         u32 type;
 273         u32 tag = blk_mq_unique_tag(req);
 274
 275         switch (req_op(req)) {
 276         case REQ_OP_DISCARD:
 277                 type = NBD_CMD_TRIM;
 278                 break;
 279         case REQ_OP_FLUSH:
 280                 type = NBD_CMD_FLUSH;
 281                 break;
 282         case REQ_OP_WRITE:
 283                 type = NBD_CMD_WRITE;
 284                 break;
 285         case REQ_OP_READ:
 286                 type = NBD_CMD_READ;
 287                 break;
 288         default:
 289                 return -EIO;
 290         }
 291
 292         if (rq_data_dir(req) == WRITE &&
 293             (nbd->flags & NBD_FLAG_READ_ONLY)) {
 294                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 295                                     "Write on read-only\n");
 296                 return -EIO;
 297         }
 298
 299         memset(&request, 0, sizeof(request));
 300         request.magic = htonl(NBD_REQUEST_MAGIC);
 301         request.type = htonl(type);
 302         if (type != NBD_CMD_FLUSH) {
 303                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 304                 request.len = htonl(size);
 305         }
 306         memcpy(request.handle, &tag, sizeof(tag));
 307
 308         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 309                 cmd, nbdcmd_to_ascii(type),
 310                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 311         result = sock_xmit(nbd, index, 1, &request, sizeof(request),
 312                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
 313         if (result <= 0) {
 314                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 315                         "Send control failed (result %d)\n", result);
 316                 return -EIO;
 317         }
 318
 319         if (type != NBD_CMD_WRITE)
 320                 return 0;
 321
 322         bio = req->bio;
 323         while (bio) {
 324                 struct bio *next = bio->bi_next;
 325                 struct bvec_iter iter;
 326                 struct bio_vec bvec;
 327
 328                 bio_for_each_segment(bvec, bio, iter) {
 329                         bool is_last = !next && bio_iter_last(bvec, iter);
 330                         int flags = is_last ? 0 : MSG_MORE;
 331
 332                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 333                                 cmd, bvec.bv_len);
 334                         result = sock_send_bvec(nbd, index, &bvec, flags);
 335                         if (result <= 0) {
 336                                 dev_err(disk_to_dev(nbd->disk),
 337                                         "Send data failed (result %d)\n",
 338                                         result);
 339                                 return -EIO;
 340                         }
 341                         /*
 342                          * The completion might already have come in,
 343                          * so break for the last one instead of letting
 344                          * the iterator do it. This prevents use-after-free
 345                          * of the bio.
 346                          */
 347                         if (is_last)
 348                                 break;
 349                 }
 350                 bio = next;
 351         }
 352         return 0;
 353 }
 354
 355 static inline int sock_recv_bvec(struct nbd_device *nbd, int index,
 356                                  struct bio_vec *bvec)
 357 {
 358         int result;
 359         void *kaddr = kmap(bvec->bv_page);
 360         result = sock_xmit(nbd, index, 0, kaddr + bvec->bv_offset,
 361                            bvec->bv_len, MSG_WAITALL);
 362         kunmap(bvec->bv_page);
 363         return result;
 364 }
 365
 366 /* NULL returned = something went wrong, inform userspace */
 367 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 368 {
 369         int result;
 370         struct nbd_reply reply;
 371         struct nbd_cmd *cmd;
 372         struct request *req = NULL;
 373         u16 hwq;
 374         u32 tag;
 375
 376         reply.magic = 0;
 377         result = sock_xmit(nbd, index, 0, &reply, sizeof(reply), MSG_WAITALL);
 378         if (result <= 0) {
 379                 if (!test_bit(NBD_DISCONNECTED, &nbd->runtime_flags) &&
 380                     !test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags))
 381                         dev_err(disk_to_dev(nbd->disk),
 382                                 "Receive control failed (result %d)\n", result);
 383                 return ERR_PTR(result);
 384         }
 385
 386         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 387                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 388                                 (unsigned long)ntohl(reply.magic));
 389                 return ERR_PTR(-EPROTO);
 390         }
 391
 392         memcpy(&tag, reply.handle, sizeof(u32));
 393
 394         hwq = blk_mq_unique_tag_to_hwq(tag);
 395         if (hwq < nbd->tag_set.nr_hw_queues)
 396                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 397                                        blk_mq_unique_tag_to_tag(tag));
 398         if (!req || !blk_mq_request_started(req)) {
 399                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 400                         tag, req);
 401                 return ERR_PTR(-ENOENT);
 402         }
 403         cmd = blk_mq_rq_to_pdu(req);
 404         if (ntohl(reply.error)) {
 405                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 406                         ntohl(reply.error));
 407                 req->errors++;
 408                 return cmd;
 409         }
 410
 411         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", cmd);
 412         if (rq_data_dir(req) != WRITE) {
 413                 struct req_iterator iter;
 414                 struct bio_vec bvec;
 415
 416                 rq_for_each_segment(bvec, req, iter) {
 417                         result = sock_recv_bvec(nbd, index, &bvec);
 418                         if (result <= 0) {
 419                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 420                                         result);
 421                                 req->errors++;
 422                                 return cmd;
 423                         }
 424                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 425                                 cmd, bvec.bv_len);
 426                 }
 427         } else {
 428                 /* See the comment in nbd_queue_rq. */
 429                 wait_for_completion(&cmd->send_complete);
 430         }
 431         return cmd;
 432 }
 433
 434 static ssize_t pid_show(struct device *dev,
 435                         struct device_attribute *attr, char *buf)
 436 {
 437         struct gendisk *disk = dev_to_disk(dev);
 438         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 439
 440         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 441 }
 442
 443 static struct device_attribute pid_attr = {
 444         .attr = { .name = "pid", .mode = S_IRUGO},
 445         .show = pid_show,
 446 };
 447
 448 struct recv_thread_args {
 449         struct work_struct work;
 450         struct nbd_device *nbd;
 451         int index;
 452 };
 453
 454 static void recv_work(struct work_struct *work)
 455 {
 456         struct recv_thread_args *args = container_of(work,
 457                                                      struct recv_thread_args,
 458                                                      work);
 459         struct nbd_device *nbd = args->nbd;
 460         struct nbd_cmd *cmd;
 461         int ret = 0;
 462
 463         BUG_ON(nbd->magic != NBD_MAGIC);
 464         while (1) {
 465                 cmd = nbd_read_stat(nbd, args->index);
 466                 if (IS_ERR(cmd)) {
 467                         ret = PTR_ERR(cmd);
 468                         break;
 469                 }
 470
 471                 nbd_end_request(cmd);
 472         }
 473
 474         /*
 475          * We got an error, shut everybody down if this wasn't the result of a
 476          * disconnect request.
 477          */
 478         if (ret && !test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags))
 479                 sock_shutdown(nbd);
 480         atomic_dec(&nbd->recv_threads);
 481         wake_up(&nbd->recv_wq);
 482 }
 483
 484 static void nbd_clear_req(struct request *req, void *data, bool reserved)
 485 {
 486         struct nbd_cmd *cmd;
 487
 488         if (!blk_mq_request_started(req))
 489                 return;
 490         cmd = blk_mq_rq_to_pdu(req);
 491         req->errors++;
 492         nbd_end_request(cmd);
 493 }
 494
 495 static void nbd_clear_que(struct nbd_device *nbd)
 496 {
 497         BUG_ON(nbd->magic != NBD_MAGIC);
 498
 499         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 500         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 501 }
 502
 503
 504 static void nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 505 {
 506         struct request *req = blk_mq_rq_from_pdu(cmd);
 507         struct nbd_device *nbd = cmd->nbd;
 508         struct nbd_sock *nsock;
 509
 510         if (index >= nbd->num_connections) {
 511                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 512                                     "Attempted send on invalid socket\n");
 513                 goto error_out;
 514         }
 515
 516         if (test_bit(NBD_DISCONNECTED, &nbd->runtime_flags)) {
 517                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 518                                     "Attempted send on closed socket\n");
 519                 goto error_out;
 520         }
 521
 522         req->errors = 0;
 523
 524         nsock = nbd->socks[index];
 525         mutex_lock(&nsock->tx_lock);
 526         if (unlikely(!nsock->sock)) {
 527                 mutex_unlock(&nsock->tx_lock);
 528                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 529                                     "Attempted send on closed socket\n");
 530                 goto error_out;
 531         }
 532
 533         if (nbd_send_cmd(nbd, cmd, index) != 0) {
 534                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 535                                     "Request send failed\n");
 536                 req->errors++;
 537                 nbd_end_request(cmd);
 538         }
 539
 540         mutex_unlock(&nsock->tx_lock);
 541
 542         return;
 543
 544 error_out:
 545         req->errors++;
 546         nbd_end_request(cmd);
 547 }
 548
 549 static int nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 550                         const struct blk_mq_queue_data *bd)
 551 {
 552         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 553
 554         /*
 555          * Since we look at the bio's to send the request over the network we
 556          * need to make sure the completion work doesn't mark this request done
 557          * before we are done doing our send.  This keeps us from dereferencing
 558          * freed data if we have particularly fast completions (ie we get the
 559          * completion before we exit sock_xmit on the last bvec) or in the case
 560          * that the server is misbehaving (or there was an error) before we're
 561          * done sending everything over the wire.
 562          */
 563         init_completion(&cmd->send_complete);
 564         blk_mq_start_request(bd->rq);
 565         nbd_handle_cmd(cmd, hctx->queue_num);
 566         complete(&cmd->send_complete);
 567
 568         return BLK_MQ_RQ_QUEUE_OK;
 569 }
 570
 571 static int nbd_add_socket(struct nbd_device *nbd, struct socket *sock)
 572 {
 573         struct nbd_sock **socks;
 574         struct nbd_sock *nsock;
 575
 576         if (!nbd->task_setup)
 577                 nbd->task_setup = current;
 578         if (nbd->task_setup != current) {
 579                 dev_err(disk_to_dev(nbd->disk),
 580                         "Device being setup by another task");
 581                 return -EINVAL;
 582         }
 583
 584         socks = krealloc(nbd->socks, (nbd->num_connections + 1) *
 585                          sizeof(struct nbd_sock *), GFP_KERNEL);
 586         if (!socks)
 587                 return -ENOMEM;
 588         nsock = kzalloc(sizeof(struct nbd_sock), GFP_KERNEL);
 589         if (!nsock)
 590                 return -ENOMEM;
 591
 592         nbd->socks = socks;
 593
 594         mutex_init(&nsock->tx_lock);
 595         nsock->sock = sock;
 596         socks[nbd->num_connections++] = nsock;
 597
 598         return 0;
 599 }
 600
 601 /* Reset all properties of an NBD device */
 602 static void nbd_reset(struct nbd_device *nbd)
 603 {
 604         int i;
 605
 606         for (i = 0; i < nbd->num_connections; i++)
 607                 kfree(nbd->socks[i]);
 608         kfree(nbd->socks);
 609         nbd->socks = NULL;
 610         nbd->runtime_flags = 0;
 611         nbd->blksize = 1024;
 612         nbd->bytesize = 0;
 613         set_capacity(nbd->disk, 0);
 614         nbd->flags = 0;
 615         nbd->tag_set.timeout = 0;
 616         nbd->num_connections = 0;
 617         nbd->task_setup = NULL;
 618         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 619 }
 620
 621 static void nbd_bdev_reset(struct block_device *bdev)
 622 {
 623         set_device_ro(bdev, false);
 624         bdev->bd_inode->i_size = 0;
 625         if (max_part > 0) {
 626                 blkdev_reread_part(bdev);
 627                 bdev->bd_invalidated = 1;
 628         }
 629 }
 630
 631 static void nbd_parse_flags(struct nbd_device *nbd, struct block_device *bdev)
 632 {
 633         if (nbd->flags & NBD_FLAG_READ_ONLY)
 634                 set_device_ro(bdev, true);
 635         if (nbd->flags & NBD_FLAG_SEND_TRIM)
 636                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 637         if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 638                 blk_queue_write_cache(nbd->disk->queue, true, false);
 639         else
 640                 blk_queue_write_cache(nbd->disk->queue, false, false);
 641 }
 642
 643 static void send_disconnects(struct nbd_device *nbd)
 644 {
 645         struct nbd_request request = {};
 646         int i, ret;
 647
 648         request.magic = htonl(NBD_REQUEST_MAGIC);
 649         request.type = htonl(NBD_CMD_DISC);
 650
 651         for (i = 0; i < nbd->num_connections; i++) {
 652                 ret = sock_xmit(nbd, i, 1, &request, sizeof(request), 0);
 653                 if (ret <= 0)
 654                         dev_err(disk_to_dev(nbd->disk),
 655                                 "Send disconnect failed %d\n", ret);
 656         }
 657 }
 658
 659 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 660 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 661
 662 /* Must be called with config_lock held */
 663 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 664                        unsigned int cmd, unsigned long arg)
 665 {
 666         switch (cmd) {
 667         case NBD_DISCONNECT: {
 668                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 669                 if (!nbd->socks)
 670                         return -EINVAL;
 671
 672                 mutex_unlock(&nbd->config_lock);
 673                 fsync_bdev(bdev);
 674                 mutex_lock(&nbd->config_lock);
 675
 676                 /* Check again after getting mutex back.  */
 677                 if (!nbd->socks)
 678                         return -EINVAL;
 679
 680                 if (!test_and_set_bit(NBD_DISCONNECT_REQUESTED,
 681                                       &nbd->runtime_flags))
 682                         send_disconnects(nbd);
 683                 return 0;
 684         }
 685
 686         case NBD_CLEAR_SOCK:
 687                 sock_shutdown(nbd);
 688                 nbd_clear_que(nbd);
 689                 kill_bdev(bdev);
 690                 nbd_bdev_reset(bdev);
 691                 /*
 692                  * We want to give the run thread a chance to wait for everybody
 693                  * to clean up and then do it's own cleanup.
 694                  */
 695                 if (!test_bit(NBD_RUNNING, &nbd->runtime_flags)) {
 696                         int i;
 697
 698                         for (i = 0; i < nbd->num_connections; i++)
 699                                 kfree(nbd->socks[i]);
 700                         kfree(nbd->socks);
 701                         nbd->socks = NULL;
 702                         nbd->num_connections = 0;
 703                         nbd->task_setup = NULL;
 704                 }
 705                 return 0;
 706
 707         case NBD_SET_SOCK: {
 708                 int err;
 709                 struct socket *sock = sockfd_lookup(arg, &err);
 710
 711                 if (!sock)
 712                         return err;
 713
 714                 err = nbd_add_socket(nbd, sock);
 715                 if (!err && max_part)
 716                         bdev->bd_invalidated = 1;
 717
 718                 return err;
 719         }
 720
 721         case NBD_SET_BLKSIZE: {
 722                 loff_t bsize = div_s64(nbd->bytesize, arg);
 723
 724                 return nbd_size_set(nbd, bdev, arg, bsize);
 725         }
 726
 727         case NBD_SET_SIZE:
 728                 return nbd_size_set(nbd, bdev, nbd->blksize,
 729                                         div_s64(arg, nbd->blksize));
 730
 731         case NBD_SET_SIZE_BLOCKS:
 732                 return nbd_size_set(nbd, bdev, nbd->blksize, arg);
 733
 734         case NBD_SET_TIMEOUT:
 735                 nbd->tag_set.timeout = arg * HZ;
 736                 return 0;
 737
 738         case NBD_SET_FLAGS:
 739                 nbd->flags = arg;
 740                 return 0;
 741
 742         case NBD_DO_IT: {
 743                 struct recv_thread_args *args;
 744                 int num_connections = nbd->num_connections;
 745                 int error = 0, i;
 746
 747                 if (nbd->task_recv)
 748                         return -EBUSY;
 749                 if (!nbd->socks)
 750                         return -EINVAL;
 751                 if (num_connections > 1 &&
 752                     !(nbd->flags & NBD_FLAG_CAN_MULTI_CONN)) {
 753                         dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
 754                         error = -EINVAL;
 755                         goto out_err;
 756                 }
 757
 758                 set_bit(NBD_RUNNING, &nbd->runtime_flags);
 759                 blk_mq_update_nr_hw_queues(&nbd->tag_set, nbd->num_connections);
 760                 args = kcalloc(num_connections, sizeof(*args), GFP_KERNEL);
 761                 if (!args) {
 762                         error = -ENOMEM;
 763                         goto out_err;
 764                 }
 765                 nbd->task_recv = current;
 766                 mutex_unlock(&nbd->config_lock);
 767
 768                 nbd_parse_flags(nbd, bdev);
 769
 770                 error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 771                 if (error) {
 772                         dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 773                         goto out_recv;
 774                 }
 775
 776                 nbd_size_update(nbd, bdev);
 777
 778                 nbd_dev_dbg_init(nbd);
 779                 for (i = 0; i < num_connections; i++) {
 780                         sk_set_memalloc(nbd->socks[i]->sock->sk);
 781                         atomic_inc(&nbd->recv_threads);
 782                         INIT_WORK(&args[i].work, recv_work);
 783                         args[i].nbd = nbd;
 784                         args[i].index = i;
 785                         queue_work(recv_workqueue, &args[i].work);
 786                 }
 787                 wait_event_interruptible(nbd->recv_wq,
 788                                          atomic_read(&nbd->recv_threads) == 0);
 789                 for (i = 0; i < num_connections; i++)
 790                         flush_work(&args[i].work);
 791                 nbd_dev_dbg_close(nbd);
 792                 nbd_size_clear(nbd, bdev);
 793                 device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 794 out_recv:
 795                 mutex_lock(&nbd->config_lock);
 796                 nbd->task_recv = NULL;
 797 out_err:
 798                 sock_shutdown(nbd);
 799                 nbd_clear_que(nbd);
 800                 kill_bdev(bdev);
 801                 nbd_bdev_reset(bdev);
 802
 803                 /* user requested, ignore socket errors */
 804                 if (test_bit(NBD_DISCONNECT_REQUESTED, &nbd->runtime_flags))
 805                         error = 0;
 806                 if (test_bit(NBD_TIMEDOUT, &nbd->runtime_flags))
 807                         error = -ETIMEDOUT;
 808
 809                 nbd_reset(nbd);
 810                 return error;
 811         }
 812
 813         case NBD_CLEAR_QUE:
 814                 /*
 815                  * This is for compatibility only.  The queue is always cleared
 816                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 817                  */
 818                 return 0;
 819
 820         case NBD_PRINT_DEBUG:
 821                 /*
 822                  * For compatibility only, we no longer keep a list of
 823                  * outstanding requests.
 824                  */
 825                 return 0;
 826         }
 827         return -ENOTTY;
 828 }
 829
 830 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 831                      unsigned int cmd, unsigned long arg)
 832 {
 833         struct nbd_device *nbd = bdev->bd_disk->private_data;
 834         int error;
 835
 836         if (!capable(CAP_SYS_ADMIN))
 837                 return -EPERM;
 838
 839         BUG_ON(nbd->magic != NBD_MAGIC);
 840
 841         mutex_lock(&nbd->config_lock);
 842         error = __nbd_ioctl(bdev, nbd, cmd, arg);
 843         mutex_unlock(&nbd->config_lock);
 844
 845         return error;
 846 }
 847
 848 static const struct block_device_operations nbd_fops =
 849 {
 850         .owner =        THIS_MODULE,
 851         .ioctl =        nbd_ioctl,
 852         .compat_ioctl = nbd_ioctl,
 853 };
 854
 855 #if IS_ENABLED(CONFIG_DEBUG_FS)
 856
 857 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
 858 {
 859         struct nbd_device *nbd = s->private;
 860
 861         if (nbd->task_recv)
 862                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
 863
 864         return 0;
 865 }
 866
 867 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
 868 {
 869         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
 870 }
 871
 872 static const struct file_operations nbd_dbg_tasks_ops = {
 873         .open = nbd_dbg_tasks_open,
 874         .read = seq_read,
 875         .llseek = seq_lseek,
 876         .release = single_release,
 877 };
 878
 879 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
 880 {
 881         struct nbd_device *nbd = s->private;
 882         u32 flags = nbd->flags;
 883
 884         seq_printf(s, "Hex: 0x%08x\n\n", flags);
 885
 886         seq_puts(s, "Known flags:\n");
 887
 888         if (flags & NBD_FLAG_HAS_FLAGS)
 889                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
 890         if (flags & NBD_FLAG_READ_ONLY)
 891                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
 892         if (flags & NBD_FLAG_SEND_FLUSH)
 893                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
 894         if (flags & NBD_FLAG_SEND_TRIM)
 895                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
 896
 897         return 0;
 898 }
 899
 900 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
 901 {
 902         return single_open(file, nbd_dbg_flags_show, inode->i_private);
 903 }
 904
 905 static const struct file_operations nbd_dbg_flags_ops = {
 906         .open = nbd_dbg_flags_open,
 907         .read = seq_read,
 908         .llseek = seq_lseek,
 909         .release = single_release,
 910 };
 911
 912 static int nbd_dev_dbg_init(struct nbd_device *nbd)
 913 {
 914         struct dentry *dir;
 915
 916         if (!nbd_dbg_dir)
 917                 return -EIO;
 918
 919         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
 920         if (!dir) {
 921                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
 922                         nbd_name(nbd));
 923                 return -EIO;
 924         }
 925         nbd->dbg_dir = dir;
 926
 927         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
 928         debugfs_create_u64("size_bytes", 0444, dir, &nbd->bytesize);
 929         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
 930         debugfs_create_u64("blocksize", 0444, dir, &nbd->blksize);
 931         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
 932
 933         return 0;
 934 }
 935
 936 static void nbd_dev_dbg_close(struct nbd_device *nbd)
 937 {
 938         debugfs_remove_recursive(nbd->dbg_dir);
 939 }
 940
 941 static int nbd_dbg_init(void)
 942 {
 943         struct dentry *dbg_dir;
 944
 945         dbg_dir = debugfs_create_dir("nbd", NULL);
 946         if (!dbg_dir)
 947                 return -EIO;
 948
 949         nbd_dbg_dir = dbg_dir;
 950
 951         return 0;
 952 }
 953
 954 static void nbd_dbg_close(void)
 955 {
 956         debugfs_remove_recursive(nbd_dbg_dir);
 957 }
 958
 959 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
 960
 961 static int nbd_dev_dbg_init(struct nbd_device *nbd)
 962 {
 963         return 0;
 964 }
 965
 966 static void nbd_dev_dbg_close(struct nbd_device *nbd)
 967 {
 968 }
 969
 970 static int nbd_dbg_init(void)
 971 {
 972         return 0;
 973 }
 974
 975 static void nbd_dbg_close(void)
 976 {
 977 }
 978
 979 #endif
 980
 981 static int nbd_init_request(void *data, struct request *rq,
 982                             unsigned int hctx_idx, unsigned int request_idx,
 983                             unsigned int numa_node)
 984 {
 985         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
 986         cmd->nbd = data;
 987         return 0;
 988 }
 989
 990 static struct blk_mq_ops nbd_mq_ops = {
 991         .queue_rq       = nbd_queue_rq,
 992         .init_request   = nbd_init_request,
 993         .timeout        = nbd_xmit_timeout,
 994 };
 995
 996 /*
 997  * And here should be modules and kernel interface
 998  *  (Just smiley confuses emacs :-)
 999  */
1000
1001 static int __init nbd_init(void)
1002 {
1003         int err = -ENOMEM;
1004         int i;
1005         int part_shift;
1006
1007         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
1008
1009         if (max_part < 0) {
1010                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
1011                 return -EINVAL;
1012         }
1013
1014         part_shift = 0;
1015         if (max_part > 0) {
1016                 part_shift = fls(max_part);
1017
1018                 /*
1019                  * Adjust max_part according to part_shift as it is exported
1020                  * to user space so that user can know the max number of
1021                  * partition kernel should be able to manage.
1022                  *
1023                  * Note that -1 is required because partition 0 is reserved
1024                  * for the whole disk.
1025                  */
1026                 max_part = (1UL << part_shift) - 1;
1027         }
1028
1029         if ((1UL << part_shift) > DISK_MAX_PARTS)
1030                 return -EINVAL;
1031
1032         if (nbds_max > 1UL << (MINORBITS - part_shift))
1033                 return -EINVAL;
1034         recv_workqueue = alloc_workqueue("knbd-recv",
1035                                          WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1036         if (!recv_workqueue)
1037                 return -ENOMEM;
1038
1039         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
1040         if (!nbd_dev) {
1041                 destroy_workqueue(recv_workqueue);
1042                 return -ENOMEM;
1043         }
1044
1045         for (i = 0; i < nbds_max; i++) {
1046                 struct request_queue *q;
1047                 struct gendisk *disk = alloc_disk(1 << part_shift);
1048                 if (!disk)
1049                         goto out;
1050                 nbd_dev[i].disk = disk;
1051
1052                 nbd_dev[i].tag_set.ops = &nbd_mq_ops;
1053                 nbd_dev[i].tag_set.nr_hw_queues = 1;
1054                 nbd_dev[i].tag_set.queue_depth = 128;
1055                 nbd_dev[i].tag_set.numa_node = NUMA_NO_NODE;
1056                 nbd_dev[i].tag_set.cmd_size = sizeof(struct nbd_cmd);
1057                 nbd_dev[i].tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1058                         BLK_MQ_F_SG_MERGE | BLK_MQ_F_BLOCKING;
1059                 nbd_dev[i].tag_set.driver_data = &nbd_dev[i];
1060
1061                 err = blk_mq_alloc_tag_set(&nbd_dev[i].tag_set);
1062                 if (err) {
1063                         put_disk(disk);
1064                         goto out;
1065                 }
1066
1067                 /*
1068                  * The new linux 2.5 block layer implementation requires
1069                  * every gendisk to have its very own request_queue struct.
1070                  * These structs are big so we dynamically allocate them.
1071                  */
1072                 q = blk_mq_init_queue(&nbd_dev[i].tag_set);
1073                 if (IS_ERR(q)) {
1074                         blk_mq_free_tag_set(&nbd_dev[i].tag_set);
1075                         put_disk(disk);
1076                         goto out;
1077                 }
1078                 disk->queue = q;
1079
1080                 /*
1081                  * Tell the block layer that we are not a rotational device
1082                  */
1083                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
1084                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1085                 disk->queue->limits.discard_granularity = 512;
1086                 blk_queue_max_discard_sectors(disk->queue, UINT_MAX);
1087                 disk->queue->limits.discard_zeroes_data = 0;
1088                 blk_queue_max_hw_sectors(disk->queue, 65536);
1089                 disk->queue->limits.max_sectors = 256;
1090         }
1091
1092         if (register_blkdev(NBD_MAJOR, "nbd")) {
1093                 err = -EIO;
1094                 goto out;
1095         }
1096
1097         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
1098
1099         nbd_dbg_init();
1100
1101         for (i = 0; i < nbds_max; i++) {
1102                 struct gendisk *disk = nbd_dev[i].disk;
1103                 nbd_dev[i].magic = NBD_MAGIC;
1104                 mutex_init(&nbd_dev[i].config_lock);
1105                 disk->major = NBD_MAJOR;
1106                 disk->first_minor = i << part_shift;
1107                 disk->fops = &nbd_fops;
1108                 disk->private_data = &nbd_dev[i];
1109                 sprintf(disk->disk_name, "nbd%d", i);
1110                 init_waitqueue_head(&nbd_dev[i].recv_wq);
1111                 nbd_reset(&nbd_dev[i]);
1112                 add_disk(disk);
1113         }
1114
1115         return 0;
1116 out:
1117         while (i--) {
1118                 blk_mq_free_tag_set(&nbd_dev[i].tag_set);
1119                 blk_cleanup_queue(nbd_dev[i].disk->queue);
1120                 put_disk(nbd_dev[i].disk);
1121         }
1122         kfree(nbd_dev);
1123         destroy_workqueue(recv_workqueue);
1124         return err;
1125 }
1126
1127 static void __exit nbd_cleanup(void)
1128 {
1129         int i;
1130
1131         nbd_dbg_close();
1132
1133         for (i = 0; i < nbds_max; i++) {
1134                 struct gendisk *disk = nbd_dev[i].disk;
1135                 nbd_dev[i].magic = 0;
1136                 if (disk) {
1137                         del_gendisk(disk);
1138                         blk_cleanup_queue(disk->queue);
1139                         blk_mq_free_tag_set(&nbd_dev[i].tag_set);
1140                         put_disk(disk);
1141                 }
1142         }
1143         destroy_workqueue(recv_workqueue);
1144         unregister_blkdev(NBD_MAJOR, "nbd");
1145         kfree(nbd_dev);
1146         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
1147 }
1148
1149 module_init(nbd_init);
1150 module_exit(nbd_cleanup);
1151
1152 MODULE_DESCRIPTION("Network Block Device");
1153 MODULE_LICENSE("GPL");
1154
1155 module_param(nbds_max, int, 0444);
1156 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
1157 module_param(max_part, int, 0444);
1158 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");