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 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/compiler.h> 28 #include <linux/err.h> 29 #include <linux/kernel.h> 30 #include <net/sock.h> 31 32 #include <asm/uaccess.h> 33 #include <asm/system.h> 34 #include <asm/types.h> 35 36 #include <linux/nbd.h> 37 38 #define LO_MAGIC 0x68797548 39 40 #ifdef NDEBUG 41 #define dprintk(flags, fmt...) 42 #else /* NDEBUG */ 43 #define dprintk(flags, fmt...) do { \ 44 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \ 45 } while (0) 46 #define DBG_IOCTL 0x0004 47 #define DBG_INIT 0x0010 48 #define DBG_EXIT 0x0020 49 #define DBG_BLKDEV 0x0100 50 #define DBG_RX 0x0200 51 #define DBG_TX 0x0400 52 static unsigned int debugflags; 53 #endif /* NDEBUG */ 54 55 static unsigned int nbds_max = 16; 56 static struct nbd_device nbd_dev[MAX_NBD]; 57 58 /* 59 * Use just one lock (or at most 1 per NIC). Two arguments for this: 60 * 1. Each NIC is essentially a synchronization point for all servers 61 * accessed through that NIC so there's no need to have more locks 62 * than NICs anyway. 63 * 2. More locks lead to more "Dirty cache line bouncing" which will slow 64 * down each lock to the point where they're actually slower than just 65 * a single lock. 66 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this! 67 */ 68 static DEFINE_SPINLOCK(nbd_lock); 69 70 #ifndef NDEBUG 71 static const char *ioctl_cmd_to_ascii(int cmd) 72 { 73 switch (cmd) { 74 case NBD_SET_SOCK: return "set-sock"; 75 case NBD_SET_BLKSIZE: return "set-blksize"; 76 case NBD_SET_SIZE: return "set-size"; 77 case NBD_DO_IT: return "do-it"; 78 case NBD_CLEAR_SOCK: return "clear-sock"; 79 case NBD_CLEAR_QUE: return "clear-que"; 80 case NBD_PRINT_DEBUG: return "print-debug"; 81 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks"; 82 case NBD_DISCONNECT: return "disconnect"; 83 case BLKROSET: return "set-read-only"; 84 case BLKFLSBUF: return "flush-buffer-cache"; 85 } 86 return "unknown"; 87 } 88 89 static const char *nbdcmd_to_ascii(int cmd) 90 { 91 switch (cmd) { 92 case NBD_CMD_READ: return "read"; 93 case NBD_CMD_WRITE: return "write"; 94 case NBD_CMD_DISC: return "disconnect"; 95 } 96 return "invalid"; 97 } 98 #endif /* NDEBUG */ 99 100 static void nbd_end_request(struct request *req) 101 { 102 int uptodate = (req->errors == 0) ? 1 : 0; 103 request_queue_t *q = req->q; 104 unsigned long flags; 105 106 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name, 107 req, uptodate? "done": "failed"); 108 109 spin_lock_irqsave(q->queue_lock, flags); 110 if (!end_that_request_first(req, uptodate, req->nr_sectors)) { 111 end_that_request_last(req, uptodate); 112 } 113 spin_unlock_irqrestore(q->queue_lock, flags); 114 } 115 116 /* 117 * Send or receive packet. 118 */ 119 static int sock_xmit(struct socket *sock, int send, void *buf, int size, 120 int msg_flags) 121 { 122 int result; 123 struct msghdr msg; 124 struct kvec iov; 125 unsigned long flags; 126 sigset_t oldset; 127 128 /* Allow interception of SIGKILL only 129 * Don't allow other signals to interrupt the transmission */ 130 spin_lock_irqsave(¤t->sighand->siglock, flags); 131 oldset = current->blocked; 132 sigfillset(¤t->blocked); 133 sigdelsetmask(¤t->blocked, sigmask(SIGKILL)); 134 recalc_sigpending(); 135 spin_unlock_irqrestore(¤t->sighand->siglock, flags); 136 137 do { 138 sock->sk->sk_allocation = GFP_NOIO; 139 iov.iov_base = buf; 140 iov.iov_len = size; 141 msg.msg_name = NULL; 142 msg.msg_namelen = 0; 143 msg.msg_control = NULL; 144 msg.msg_controllen = 0; 145 msg.msg_flags = msg_flags | MSG_NOSIGNAL; 146 147 if (send) 148 result = kernel_sendmsg(sock, &msg, &iov, 1, size); 149 else 150 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0); 151 152 if (signal_pending(current)) { 153 siginfo_t info; 154 spin_lock_irqsave(¤t->sighand->siglock, flags); 155 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n", 156 current->pid, current->comm, 157 dequeue_signal(current, ¤t->blocked, &info)); 158 spin_unlock_irqrestore(¤t->sighand->siglock, flags); 159 result = -EINTR; 160 break; 161 } 162 163 if (result <= 0) { 164 if (result == 0) 165 result = -EPIPE; /* short read */ 166 break; 167 } 168 size -= result; 169 buf += result; 170 } while (size > 0); 171 172 spin_lock_irqsave(¤t->sighand->siglock, flags); 173 current->blocked = oldset; 174 recalc_sigpending(); 175 spin_unlock_irqrestore(¤t->sighand->siglock, flags); 176 177 return result; 178 } 179 180 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec, 181 int flags) 182 { 183 int result; 184 void *kaddr = kmap(bvec->bv_page); 185 result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len, 186 flags); 187 kunmap(bvec->bv_page); 188 return result; 189 } 190 191 static int nbd_send_req(struct nbd_device *lo, struct request *req) 192 { 193 int result, i, flags; 194 struct nbd_request request; 195 unsigned long size = req->nr_sectors << 9; 196 struct socket *sock = lo->sock; 197 198 request.magic = htonl(NBD_REQUEST_MAGIC); 199 request.type = htonl(nbd_cmd(req)); 200 request.from = cpu_to_be64((u64) req->sector << 9); 201 request.len = htonl(size); 202 memcpy(request.handle, &req, sizeof(req)); 203 204 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n", 205 lo->disk->disk_name, req, 206 nbdcmd_to_ascii(nbd_cmd(req)), 207 (unsigned long long)req->sector << 9, 208 req->nr_sectors << 9); 209 result = sock_xmit(sock, 1, &request, sizeof(request), 210 (nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0); 211 if (result <= 0) { 212 printk(KERN_ERR "%s: Send control failed (result %d)\n", 213 lo->disk->disk_name, result); 214 goto error_out; 215 } 216 217 if (nbd_cmd(req) == NBD_CMD_WRITE) { 218 struct bio *bio; 219 /* 220 * we are really probing at internals to determine 221 * whether to set MSG_MORE or not... 222 */ 223 rq_for_each_bio(bio, req) { 224 struct bio_vec *bvec; 225 bio_for_each_segment(bvec, bio, i) { 226 flags = 0; 227 if ((i < (bio->bi_vcnt - 1)) || bio->bi_next) 228 flags = MSG_MORE; 229 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n", 230 lo->disk->disk_name, req, 231 bvec->bv_len); 232 result = sock_send_bvec(sock, bvec, flags); 233 if (result <= 0) { 234 printk(KERN_ERR "%s: Send data failed (result %d)\n", 235 lo->disk->disk_name, 236 result); 237 goto error_out; 238 } 239 } 240 } 241 } 242 return 0; 243 244 error_out: 245 return 1; 246 } 247 248 static struct request *nbd_find_request(struct nbd_device *lo, char *handle) 249 { 250 struct request *req; 251 struct list_head *tmp; 252 struct request *xreq; 253 int err; 254 255 memcpy(&xreq, handle, sizeof(xreq)); 256 257 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq); 258 if (unlikely(err)) 259 goto out; 260 261 spin_lock(&lo->queue_lock); 262 list_for_each(tmp, &lo->queue_head) { 263 req = list_entry(tmp, struct request, queuelist); 264 if (req != xreq) 265 continue; 266 list_del_init(&req->queuelist); 267 spin_unlock(&lo->queue_lock); 268 return req; 269 } 270 spin_unlock(&lo->queue_lock); 271 272 err = -ENOENT; 273 274 out: 275 return ERR_PTR(err); 276 } 277 278 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec) 279 { 280 int result; 281 void *kaddr = kmap(bvec->bv_page); 282 result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len, 283 MSG_WAITALL); 284 kunmap(bvec->bv_page); 285 return result; 286 } 287 288 /* NULL returned = something went wrong, inform userspace */ 289 static struct request *nbd_read_stat(struct nbd_device *lo) 290 { 291 int result; 292 struct nbd_reply reply; 293 struct request *req; 294 struct socket *sock = lo->sock; 295 296 reply.magic = 0; 297 result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL); 298 if (result <= 0) { 299 printk(KERN_ERR "%s: Receive control failed (result %d)\n", 300 lo->disk->disk_name, result); 301 goto harderror; 302 } 303 req = nbd_find_request(lo, reply.handle); 304 if (unlikely(IS_ERR(req))) { 305 result = PTR_ERR(req); 306 if (result != -ENOENT) 307 goto harderror; 308 309 printk(KERN_ERR "%s: Unexpected reply (%p)\n", 310 lo->disk->disk_name, reply.handle); 311 result = -EBADR; 312 goto harderror; 313 } 314 315 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) { 316 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n", 317 lo->disk->disk_name, 318 (unsigned long)ntohl(reply.magic)); 319 result = -EPROTO; 320 goto harderror; 321 } 322 if (ntohl(reply.error)) { 323 printk(KERN_ERR "%s: Other side returned error (%d)\n", 324 lo->disk->disk_name, ntohl(reply.error)); 325 req->errors++; 326 return req; 327 } 328 329 dprintk(DBG_RX, "%s: request %p: got reply\n", 330 lo->disk->disk_name, req); 331 if (nbd_cmd(req) == NBD_CMD_READ) { 332 int i; 333 struct bio *bio; 334 rq_for_each_bio(bio, req) { 335 struct bio_vec *bvec; 336 bio_for_each_segment(bvec, bio, i) { 337 result = sock_recv_bvec(sock, bvec); 338 if (result <= 0) { 339 printk(KERN_ERR "%s: Receive data failed (result %d)\n", 340 lo->disk->disk_name, 341 result); 342 goto harderror; 343 } 344 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n", 345 lo->disk->disk_name, req, bvec->bv_len); 346 } 347 } 348 } 349 return req; 350 harderror: 351 lo->harderror = result; 352 return NULL; 353 } 354 355 static void nbd_do_it(struct nbd_device *lo) 356 { 357 struct request *req; 358 359 BUG_ON(lo->magic != LO_MAGIC); 360 361 while ((req = nbd_read_stat(lo)) != NULL) 362 nbd_end_request(req); 363 return; 364 } 365 366 static void nbd_clear_que(struct nbd_device *lo) 367 { 368 struct request *req; 369 370 BUG_ON(lo->magic != LO_MAGIC); 371 372 /* 373 * Because we have set lo->sock to NULL under the tx_lock, all 374 * modifications to the list must have completed by now. For 375 * the same reason, the active_req must be NULL. 376 * 377 * As a consequence, we don't need to take the spin lock while 378 * purging the list here. 379 */ 380 BUG_ON(lo->sock); 381 BUG_ON(lo->active_req); 382 383 while (!list_empty(&lo->queue_head)) { 384 req = list_entry(lo->queue_head.next, struct request, 385 queuelist); 386 list_del_init(&req->queuelist); 387 req->errors++; 388 nbd_end_request(req); 389 } 390 } 391 392 /* 393 * We always wait for result of write, for now. It would be nice to make it optional 394 * in future 395 * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK)) 396 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); } 397 */ 398 399 static void do_nbd_request(request_queue_t * q) 400 { 401 struct request *req; 402 403 while ((req = elv_next_request(q)) != NULL) { 404 struct nbd_device *lo; 405 406 blkdev_dequeue_request(req); 407 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%lx)\n", 408 req->rq_disk->disk_name, req, req->flags); 409 410 if (!(req->flags & REQ_CMD)) 411 goto error_out; 412 413 lo = req->rq_disk->private_data; 414 415 BUG_ON(lo->magic != LO_MAGIC); 416 417 nbd_cmd(req) = NBD_CMD_READ; 418 if (rq_data_dir(req) == WRITE) { 419 nbd_cmd(req) = NBD_CMD_WRITE; 420 if (lo->flags & NBD_READ_ONLY) { 421 printk(KERN_ERR "%s: Write on read-only\n", 422 lo->disk->disk_name); 423 goto error_out; 424 } 425 } 426 427 req->errors = 0; 428 spin_unlock_irq(q->queue_lock); 429 430 mutex_lock(&lo->tx_lock); 431 if (unlikely(!lo->sock)) { 432 mutex_unlock(&lo->tx_lock); 433 printk(KERN_ERR "%s: Attempted send on closed socket\n", 434 lo->disk->disk_name); 435 req->errors++; 436 nbd_end_request(req); 437 spin_lock_irq(q->queue_lock); 438 continue; 439 } 440 441 lo->active_req = req; 442 443 if (nbd_send_req(lo, req) != 0) { 444 printk(KERN_ERR "%s: Request send failed\n", 445 lo->disk->disk_name); 446 req->errors++; 447 nbd_end_request(req); 448 } else { 449 spin_lock(&lo->queue_lock); 450 list_add(&req->queuelist, &lo->queue_head); 451 spin_unlock(&lo->queue_lock); 452 } 453 454 lo->active_req = NULL; 455 mutex_unlock(&lo->tx_lock); 456 wake_up_all(&lo->active_wq); 457 458 spin_lock_irq(q->queue_lock); 459 continue; 460 461 error_out: 462 req->errors++; 463 spin_unlock(q->queue_lock); 464 nbd_end_request(req); 465 spin_lock(q->queue_lock); 466 } 467 return; 468 } 469 470 static int nbd_ioctl(struct inode *inode, struct file *file, 471 unsigned int cmd, unsigned long arg) 472 { 473 struct nbd_device *lo = inode->i_bdev->bd_disk->private_data; 474 int error; 475 struct request sreq ; 476 477 if (!capable(CAP_SYS_ADMIN)) 478 return -EPERM; 479 480 BUG_ON(lo->magic != LO_MAGIC); 481 482 /* Anyone capable of this syscall can do *real bad* things */ 483 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n", 484 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg); 485 486 switch (cmd) { 487 case NBD_DISCONNECT: 488 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name); 489 sreq.flags = REQ_SPECIAL; 490 nbd_cmd(&sreq) = NBD_CMD_DISC; 491 /* 492 * Set these to sane values in case server implementation 493 * fails to check the request type first and also to keep 494 * debugging output cleaner. 495 */ 496 sreq.sector = 0; 497 sreq.nr_sectors = 0; 498 if (!lo->sock) 499 return -EINVAL; 500 nbd_send_req(lo, &sreq); 501 return 0; 502 503 case NBD_CLEAR_SOCK: 504 error = 0; 505 mutex_lock(&lo->tx_lock); 506 lo->sock = NULL; 507 mutex_unlock(&lo->tx_lock); 508 file = lo->file; 509 lo->file = NULL; 510 nbd_clear_que(lo); 511 BUG_ON(!list_empty(&lo->queue_head)); 512 if (file) 513 fput(file); 514 return error; 515 case NBD_SET_SOCK: 516 if (lo->file) 517 return -EBUSY; 518 error = -EINVAL; 519 file = fget(arg); 520 if (file) { 521 inode = file->f_dentry->d_inode; 522 if (S_ISSOCK(inode->i_mode)) { 523 lo->file = file; 524 lo->sock = SOCKET_I(inode); 525 error = 0; 526 } else { 527 fput(file); 528 } 529 } 530 return error; 531 case NBD_SET_BLKSIZE: 532 lo->blksize = arg; 533 lo->bytesize &= ~(lo->blksize-1); 534 inode->i_bdev->bd_inode->i_size = lo->bytesize; 535 set_blocksize(inode->i_bdev, lo->blksize); 536 set_capacity(lo->disk, lo->bytesize >> 9); 537 return 0; 538 case NBD_SET_SIZE: 539 lo->bytesize = arg & ~(lo->blksize-1); 540 inode->i_bdev->bd_inode->i_size = lo->bytesize; 541 set_blocksize(inode->i_bdev, lo->blksize); 542 set_capacity(lo->disk, lo->bytesize >> 9); 543 return 0; 544 case NBD_SET_SIZE_BLOCKS: 545 lo->bytesize = ((u64) arg) * lo->blksize; 546 inode->i_bdev->bd_inode->i_size = lo->bytesize; 547 set_blocksize(inode->i_bdev, lo->blksize); 548 set_capacity(lo->disk, lo->bytesize >> 9); 549 return 0; 550 case NBD_DO_IT: 551 if (!lo->file) 552 return -EINVAL; 553 nbd_do_it(lo); 554 /* on return tidy up in case we have a signal */ 555 /* Forcibly shutdown the socket causing all listeners 556 * to error 557 * 558 * FIXME: This code is duplicated from sys_shutdown, but 559 * there should be a more generic interface rather than 560 * calling socket ops directly here */ 561 mutex_lock(&lo->tx_lock); 562 if (lo->sock) { 563 printk(KERN_WARNING "%s: shutting down socket\n", 564 lo->disk->disk_name); 565 lo->sock->ops->shutdown(lo->sock, 566 SEND_SHUTDOWN|RCV_SHUTDOWN); 567 lo->sock = NULL; 568 } 569 mutex_unlock(&lo->tx_lock); 570 file = lo->file; 571 lo->file = NULL; 572 nbd_clear_que(lo); 573 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name); 574 if (file) 575 fput(file); 576 return lo->harderror; 577 case NBD_CLEAR_QUE: 578 /* 579 * This is for compatibility only. The queue is always cleared 580 * by NBD_DO_IT or NBD_CLEAR_SOCK. 581 */ 582 BUG_ON(!lo->sock && !list_empty(&lo->queue_head)); 583 return 0; 584 case NBD_PRINT_DEBUG: 585 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n", 586 inode->i_bdev->bd_disk->disk_name, 587 lo->queue_head.next, lo->queue_head.prev, 588 &lo->queue_head); 589 return 0; 590 } 591 return -EINVAL; 592 } 593 594 static struct block_device_operations nbd_fops = 595 { 596 .owner = THIS_MODULE, 597 .ioctl = nbd_ioctl, 598 }; 599 600 /* 601 * And here should be modules and kernel interface 602 * (Just smiley confuses emacs :-) 603 */ 604 605 static int __init nbd_init(void) 606 { 607 int err = -ENOMEM; 608 int i; 609 610 BUILD_BUG_ON(sizeof(struct nbd_request) != 28); 611 612 if (nbds_max > MAX_NBD) { 613 printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD, 614 nbds_max); 615 return -EINVAL; 616 } 617 618 for (i = 0; i < nbds_max; i++) { 619 struct gendisk *disk = alloc_disk(1); 620 if (!disk) 621 goto out; 622 nbd_dev[i].disk = disk; 623 /* 624 * The new linux 2.5 block layer implementation requires 625 * every gendisk to have its very own request_queue struct. 626 * These structs are big so we dynamically allocate them. 627 */ 628 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock); 629 if (!disk->queue) { 630 put_disk(disk); 631 goto out; 632 } 633 } 634 635 if (register_blkdev(NBD_MAJOR, "nbd")) { 636 err = -EIO; 637 goto out; 638 } 639 640 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR); 641 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags); 642 643 for (i = 0; i < nbds_max; i++) { 644 struct gendisk *disk = nbd_dev[i].disk; 645 nbd_dev[i].file = NULL; 646 nbd_dev[i].magic = LO_MAGIC; 647 nbd_dev[i].flags = 0; 648 spin_lock_init(&nbd_dev[i].queue_lock); 649 INIT_LIST_HEAD(&nbd_dev[i].queue_head); 650 mutex_init(&nbd_dev[i].tx_lock); 651 init_waitqueue_head(&nbd_dev[i].active_wq); 652 nbd_dev[i].blksize = 1024; 653 nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */ 654 disk->major = NBD_MAJOR; 655 disk->first_minor = i; 656 disk->fops = &nbd_fops; 657 disk->private_data = &nbd_dev[i]; 658 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO; 659 sprintf(disk->disk_name, "nbd%d", i); 660 set_capacity(disk, 0x7ffffc00ULL << 1); /* 2 TB */ 661 add_disk(disk); 662 } 663 664 return 0; 665 out: 666 while (i--) { 667 blk_cleanup_queue(nbd_dev[i].disk->queue); 668 put_disk(nbd_dev[i].disk); 669 } 670 return err; 671 } 672 673 static void __exit nbd_cleanup(void) 674 { 675 int i; 676 for (i = 0; i < nbds_max; i++) { 677 struct gendisk *disk = nbd_dev[i].disk; 678 nbd_dev[i].magic = 0; 679 if (disk) { 680 del_gendisk(disk); 681 blk_cleanup_queue(disk->queue); 682 put_disk(disk); 683 } 684 } 685 unregister_blkdev(NBD_MAJOR, "nbd"); 686 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR); 687 } 688 689 module_init(nbd_init); 690 module_exit(nbd_cleanup); 691 692 MODULE_DESCRIPTION("Network Block Device"); 693 MODULE_LICENSE("GPL"); 694 695 module_param(nbds_max, int, 0444); 696 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize."); 697 #ifndef NDEBUG 698 module_param(debugflags, int, 0644); 699 MODULE_PARM_DESC(debugflags, "flags for controlling debug output"); 700 #endif 701