xref: /openbmc/linux/drivers/block/nbd.c (revision d5cb9783536a41df9f9cba5b0a1d78047ed787f7)
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  * (part of code stolen from loop.c)
11  *
12  * 97-3-25 compiled 0-th version, not yet tested it
13  *   (it did not work, BTW) (later that day) HEY! it works!
14  *   (bit later) hmm, not that much... 2:00am next day:
15  *   yes, it works, but it gives something like 50kB/sec
16  * 97-4-01 complete rewrite to make it possible for many requests at
17  *   once to be processed
18  * 97-4-11 Making protocol independent of endianity etc.
19  * 97-9-13 Cosmetic changes
20  * 98-5-13 Attempt to make 64-bit-clean on 64-bit machines
21  * 99-1-11 Attempt to make 64-bit-clean on 32-bit machines <ankry@mif.pg.gda.pl>
22  * 01-2-27 Fix to store proper blockcount for kernel (calculated using
23  *   BLOCK_SIZE_BITS, not device blocksize) <aga@permonline.ru>
24  * 01-3-11 Make nbd work with new Linux block layer code. It now supports
25  *   plugging like all the other block devices. Also added in MSG_MORE to
26  *   reduce number of partial TCP segments sent. <steve@chygwyn.com>
27  * 01-12-6 Fix deadlock condition by making queue locks independent of
28  *   the transmit lock. <steve@chygwyn.com>
29  * 02-10-11 Allow hung xmit to be aborted via SIGKILL & various fixes.
30  *   <Paul.Clements@SteelEye.com> <James.Bottomley@SteelEye.com>
31  * 03-06-22 Make nbd work with new linux 2.5 block layer design. This fixes
32  *   memory corruption from module removal and possible memory corruption
33  *   from sending/receiving disk data. <ldl@aros.net>
34  * 03-06-23 Cosmetic changes. <ldl@aros.net>
35  * 03-06-23 Enhance diagnostics support. <ldl@aros.net>
36  * 03-06-24 Remove unneeded blksize_bits field from nbd_device struct.
37  *   <ldl@aros.net>
38  * 03-06-24 Cleanup PARANOIA usage & code. <ldl@aros.net>
39  * 04-02-19 Remove PARANOIA, plus various cleanups (Paul Clements)
40  * possible FIXME: make set_sock / set_blksize / set_size / do_it one syscall
41  * why not: would need access_ok and friends, would share yet another
42  *          structure with userland
43  */
44 
45 #include <linux/major.h>
46 
47 #include <linux/blkdev.h>
48 #include <linux/module.h>
49 #include <linux/init.h>
50 #include <linux/sched.h>
51 #include <linux/fs.h>
52 #include <linux/bio.h>
53 #include <linux/stat.h>
54 #include <linux/errno.h>
55 #include <linux/file.h>
56 #include <linux/ioctl.h>
57 #include <net/sock.h>
58 
59 #include <linux/devfs_fs_kernel.h>
60 
61 #include <asm/uaccess.h>
62 #include <asm/types.h>
63 
64 #include <linux/nbd.h>
65 
66 #define LO_MAGIC 0x68797548
67 
68 #ifdef NDEBUG
69 #define dprintk(flags, fmt...)
70 #else /* NDEBUG */
71 #define dprintk(flags, fmt...) do { \
72 	if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
73 } while (0)
74 #define DBG_IOCTL       0x0004
75 #define DBG_INIT        0x0010
76 #define DBG_EXIT        0x0020
77 #define DBG_BLKDEV      0x0100
78 #define DBG_RX          0x0200
79 #define DBG_TX          0x0400
80 static unsigned int debugflags;
81 static unsigned int nbds_max = 16;
82 #endif /* NDEBUG */
83 
84 static struct nbd_device nbd_dev[MAX_NBD];
85 
86 /*
87  * Use just one lock (or at most 1 per NIC). Two arguments for this:
88  * 1. Each NIC is essentially a synchronization point for all servers
89  *    accessed through that NIC so there's no need to have more locks
90  *    than NICs anyway.
91  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
92  *    down each lock to the point where they're actually slower than just
93  *    a single lock.
94  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
95  */
96 static DEFINE_SPINLOCK(nbd_lock);
97 
98 #ifndef NDEBUG
99 static const char *ioctl_cmd_to_ascii(int cmd)
100 {
101 	switch (cmd) {
102 	case NBD_SET_SOCK: return "set-sock";
103 	case NBD_SET_BLKSIZE: return "set-blksize";
104 	case NBD_SET_SIZE: return "set-size";
105 	case NBD_DO_IT: return "do-it";
106 	case NBD_CLEAR_SOCK: return "clear-sock";
107 	case NBD_CLEAR_QUE: return "clear-que";
108 	case NBD_PRINT_DEBUG: return "print-debug";
109 	case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
110 	case NBD_DISCONNECT: return "disconnect";
111 	case BLKROSET: return "set-read-only";
112 	case BLKFLSBUF: return "flush-buffer-cache";
113 	}
114 	return "unknown";
115 }
116 
117 static const char *nbdcmd_to_ascii(int cmd)
118 {
119 	switch (cmd) {
120 	case  NBD_CMD_READ: return "read";
121 	case NBD_CMD_WRITE: return "write";
122 	case  NBD_CMD_DISC: return "disconnect";
123 	}
124 	return "invalid";
125 }
126 #endif /* NDEBUG */
127 
128 static void nbd_end_request(struct request *req)
129 {
130 	int uptodate = (req->errors == 0) ? 1 : 0;
131 	request_queue_t *q = req->q;
132 	unsigned long flags;
133 
134 	dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
135 			req, uptodate? "done": "failed");
136 
137 	spin_lock_irqsave(q->queue_lock, flags);
138 	if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
139 		end_that_request_last(req);
140 	}
141 	spin_unlock_irqrestore(q->queue_lock, flags);
142 }
143 
144 /*
145  *  Send or receive packet.
146  */
147 static int sock_xmit(struct socket *sock, int send, void *buf, int size,
148 		int msg_flags)
149 {
150 	int result;
151 	struct msghdr msg;
152 	struct kvec iov;
153 	unsigned long flags;
154 	sigset_t oldset;
155 
156 	/* Allow interception of SIGKILL only
157 	 * Don't allow other signals to interrupt the transmission */
158 	spin_lock_irqsave(&current->sighand->siglock, flags);
159 	oldset = current->blocked;
160 	sigfillset(&current->blocked);
161 	sigdelsetmask(&current->blocked, sigmask(SIGKILL));
162 	recalc_sigpending();
163 	spin_unlock_irqrestore(&current->sighand->siglock, flags);
164 
165 	do {
166 		sock->sk->sk_allocation = GFP_NOIO;
167 		iov.iov_base = buf;
168 		iov.iov_len = size;
169 		msg.msg_name = NULL;
170 		msg.msg_namelen = 0;
171 		msg.msg_control = NULL;
172 		msg.msg_controllen = 0;
173 		msg.msg_namelen = 0;
174 		msg.msg_flags = msg_flags | MSG_NOSIGNAL;
175 
176 		if (send)
177 			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
178 		else
179 			result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
180 
181 		if (signal_pending(current)) {
182 			siginfo_t info;
183 			spin_lock_irqsave(&current->sighand->siglock, flags);
184 			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
185 				current->pid, current->comm,
186 				dequeue_signal(current, &current->blocked, &info));
187 			spin_unlock_irqrestore(&current->sighand->siglock, flags);
188 			result = -EINTR;
189 			break;
190 		}
191 
192 		if (result <= 0) {
193 			if (result == 0)
194 				result = -EPIPE; /* short read */
195 			break;
196 		}
197 		size -= result;
198 		buf += result;
199 	} while (size > 0);
200 
201 	spin_lock_irqsave(&current->sighand->siglock, flags);
202 	current->blocked = oldset;
203 	recalc_sigpending();
204 	spin_unlock_irqrestore(&current->sighand->siglock, flags);
205 
206 	return result;
207 }
208 
209 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
210 		int flags)
211 {
212 	int result;
213 	void *kaddr = kmap(bvec->bv_page);
214 	result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len,
215 			flags);
216 	kunmap(bvec->bv_page);
217 	return result;
218 }
219 
220 static int nbd_send_req(struct nbd_device *lo, struct request *req)
221 {
222 	int result, i, flags;
223 	struct nbd_request request;
224 	unsigned long size = req->nr_sectors << 9;
225 	struct socket *sock = lo->sock;
226 
227 	request.magic = htonl(NBD_REQUEST_MAGIC);
228 	request.type = htonl(nbd_cmd(req));
229 	request.from = cpu_to_be64((u64) req->sector << 9);
230 	request.len = htonl(size);
231 	memcpy(request.handle, &req, sizeof(req));
232 
233 	down(&lo->tx_lock);
234 
235 	if (!sock || !lo->sock) {
236 		printk(KERN_ERR "%s: Attempted send on closed socket\n",
237 				lo->disk->disk_name);
238 		goto error_out;
239 	}
240 
241 	dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
242 			lo->disk->disk_name, req,
243 			nbdcmd_to_ascii(nbd_cmd(req)),
244 			(unsigned long long)req->sector << 9,
245 			req->nr_sectors << 9);
246 	result = sock_xmit(sock, 1, &request, sizeof(request),
247 			(nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0);
248 	if (result <= 0) {
249 		printk(KERN_ERR "%s: Send control failed (result %d)\n",
250 				lo->disk->disk_name, result);
251 		goto error_out;
252 	}
253 
254 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
255 		struct bio *bio;
256 		/*
257 		 * we are really probing at internals to determine
258 		 * whether to set MSG_MORE or not...
259 		 */
260 		rq_for_each_bio(bio, req) {
261 			struct bio_vec *bvec;
262 			bio_for_each_segment(bvec, bio, i) {
263 				flags = 0;
264 				if ((i < (bio->bi_vcnt - 1)) || bio->bi_next)
265 					flags = MSG_MORE;
266 				dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
267 						lo->disk->disk_name, req,
268 						bvec->bv_len);
269 				result = sock_send_bvec(sock, bvec, flags);
270 				if (result <= 0) {
271 					printk(KERN_ERR "%s: Send data failed (result %d)\n",
272 							lo->disk->disk_name,
273 							result);
274 					goto error_out;
275 				}
276 			}
277 		}
278 	}
279 	up(&lo->tx_lock);
280 	return 0;
281 
282 error_out:
283 	up(&lo->tx_lock);
284 	return 1;
285 }
286 
287 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
288 {
289 	struct request *req;
290 	struct list_head *tmp;
291 	struct request *xreq;
292 
293 	memcpy(&xreq, handle, sizeof(xreq));
294 
295 	spin_lock(&lo->queue_lock);
296 	list_for_each(tmp, &lo->queue_head) {
297 		req = list_entry(tmp, struct request, queuelist);
298 		if (req != xreq)
299 			continue;
300 		list_del_init(&req->queuelist);
301 		spin_unlock(&lo->queue_lock);
302 		return req;
303 	}
304 	spin_unlock(&lo->queue_lock);
305 	return NULL;
306 }
307 
308 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
309 {
310 	int result;
311 	void *kaddr = kmap(bvec->bv_page);
312 	result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len,
313 			MSG_WAITALL);
314 	kunmap(bvec->bv_page);
315 	return result;
316 }
317 
318 /* NULL returned = something went wrong, inform userspace */
319 static struct request *nbd_read_stat(struct nbd_device *lo)
320 {
321 	int result;
322 	struct nbd_reply reply;
323 	struct request *req;
324 	struct socket *sock = lo->sock;
325 
326 	reply.magic = 0;
327 	result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL);
328 	if (result <= 0) {
329 		printk(KERN_ERR "%s: Receive control failed (result %d)\n",
330 				lo->disk->disk_name, result);
331 		goto harderror;
332 	}
333 	req = nbd_find_request(lo, reply.handle);
334 	if (req == NULL) {
335 		printk(KERN_ERR "%s: Unexpected reply (%p)\n",
336 				lo->disk->disk_name, reply.handle);
337 		result = -EBADR;
338 		goto harderror;
339 	}
340 
341 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
342 		printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
343 				lo->disk->disk_name,
344 				(unsigned long)ntohl(reply.magic));
345 		result = -EPROTO;
346 		goto harderror;
347 	}
348 	if (ntohl(reply.error)) {
349 		printk(KERN_ERR "%s: Other side returned error (%d)\n",
350 				lo->disk->disk_name, ntohl(reply.error));
351 		req->errors++;
352 		return req;
353 	}
354 
355 	dprintk(DBG_RX, "%s: request %p: got reply\n",
356 			lo->disk->disk_name, req);
357 	if (nbd_cmd(req) == NBD_CMD_READ) {
358 		int i;
359 		struct bio *bio;
360 		rq_for_each_bio(bio, req) {
361 			struct bio_vec *bvec;
362 			bio_for_each_segment(bvec, bio, i) {
363 				result = sock_recv_bvec(sock, bvec);
364 				if (result <= 0) {
365 					printk(KERN_ERR "%s: Receive data failed (result %d)\n",
366 							lo->disk->disk_name,
367 							result);
368 					goto harderror;
369 				}
370 				dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
371 					lo->disk->disk_name, req, bvec->bv_len);
372 			}
373 		}
374 	}
375 	return req;
376 harderror:
377 	lo->harderror = result;
378 	return NULL;
379 }
380 
381 static void nbd_do_it(struct nbd_device *lo)
382 {
383 	struct request *req;
384 
385 	BUG_ON(lo->magic != LO_MAGIC);
386 
387 	while ((req = nbd_read_stat(lo)) != NULL)
388 		nbd_end_request(req);
389 	return;
390 }
391 
392 static void nbd_clear_que(struct nbd_device *lo)
393 {
394 	struct request *req;
395 
396 	BUG_ON(lo->magic != LO_MAGIC);
397 
398 	do {
399 		req = NULL;
400 		spin_lock(&lo->queue_lock);
401 		if (!list_empty(&lo->queue_head)) {
402 			req = list_entry(lo->queue_head.next, struct request, queuelist);
403 			list_del_init(&req->queuelist);
404 		}
405 		spin_unlock(&lo->queue_lock);
406 		if (req) {
407 			req->errors++;
408 			nbd_end_request(req);
409 		}
410 	} while (req);
411 }
412 
413 /*
414  * We always wait for result of write, for now. It would be nice to make it optional
415  * in future
416  * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
417  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
418  */
419 
420 static void do_nbd_request(request_queue_t * q)
421 {
422 	struct request *req;
423 
424 	while ((req = elv_next_request(q)) != NULL) {
425 		struct nbd_device *lo;
426 
427 		blkdev_dequeue_request(req);
428 		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%lx)\n",
429 				req->rq_disk->disk_name, req, req->flags);
430 
431 		if (!(req->flags & REQ_CMD))
432 			goto error_out;
433 
434 		lo = req->rq_disk->private_data;
435 
436 		BUG_ON(lo->magic != LO_MAGIC);
437 
438 		if (!lo->file) {
439 			printk(KERN_ERR "%s: Request when not-ready\n",
440 					lo->disk->disk_name);
441 			goto error_out;
442 		}
443 		nbd_cmd(req) = NBD_CMD_READ;
444 		if (rq_data_dir(req) == WRITE) {
445 			nbd_cmd(req) = NBD_CMD_WRITE;
446 			if (lo->flags & NBD_READ_ONLY) {
447 				printk(KERN_ERR "%s: Write on read-only\n",
448 						lo->disk->disk_name);
449 				goto error_out;
450 			}
451 		}
452 
453 		req->errors = 0;
454 		spin_unlock_irq(q->queue_lock);
455 
456 		spin_lock(&lo->queue_lock);
457 
458 		if (!lo->file) {
459 			spin_unlock(&lo->queue_lock);
460 			printk(KERN_ERR "%s: failed between accept and semaphore, file lost\n",
461 					lo->disk->disk_name);
462 			req->errors++;
463 			nbd_end_request(req);
464 			spin_lock_irq(q->queue_lock);
465 			continue;
466 		}
467 
468 		list_add(&req->queuelist, &lo->queue_head);
469 		spin_unlock(&lo->queue_lock);
470 
471 		if (nbd_send_req(lo, req) != 0) {
472 			printk(KERN_ERR "%s: Request send failed\n",
473 					lo->disk->disk_name);
474 			if (nbd_find_request(lo, (char *)&req) != NULL) {
475 				/* we still own req */
476 				req->errors++;
477 				nbd_end_request(req);
478 			} else /* we're racing with nbd_clear_que */
479 				printk(KERN_DEBUG "nbd: can't find req\n");
480 		}
481 
482 		spin_lock_irq(q->queue_lock);
483 		continue;
484 
485 error_out:
486 		req->errors++;
487 		spin_unlock(q->queue_lock);
488 		nbd_end_request(req);
489 		spin_lock(q->queue_lock);
490 	}
491 	return;
492 }
493 
494 static int nbd_ioctl(struct inode *inode, struct file *file,
495 		     unsigned int cmd, unsigned long arg)
496 {
497 	struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
498 	int error;
499 	struct request sreq ;
500 
501 	if (!capable(CAP_SYS_ADMIN))
502 		return -EPERM;
503 
504 	BUG_ON(lo->magic != LO_MAGIC);
505 
506 	/* Anyone capable of this syscall can do *real bad* things */
507 	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
508 			lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
509 
510 	switch (cmd) {
511 	case NBD_DISCONNECT:
512 	        printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
513 		sreq.flags = REQ_SPECIAL;
514 		nbd_cmd(&sreq) = NBD_CMD_DISC;
515 		/*
516 		 * Set these to sane values in case server implementation
517 		 * fails to check the request type first and also to keep
518 		 * debugging output cleaner.
519 		 */
520 		sreq.sector = 0;
521 		sreq.nr_sectors = 0;
522                 if (!lo->sock)
523 			return -EINVAL;
524                 nbd_send_req(lo, &sreq);
525                 return 0;
526 
527 	case NBD_CLEAR_SOCK:
528 		error = 0;
529 		down(&lo->tx_lock);
530 		lo->sock = NULL;
531 		up(&lo->tx_lock);
532 		spin_lock(&lo->queue_lock);
533 		file = lo->file;
534 		lo->file = NULL;
535 		spin_unlock(&lo->queue_lock);
536 		nbd_clear_que(lo);
537 		spin_lock(&lo->queue_lock);
538 		if (!list_empty(&lo->queue_head)) {
539 			printk(KERN_ERR "nbd: disconnect: some requests are in progress -> please try again.\n");
540 			error = -EBUSY;
541 		}
542 		spin_unlock(&lo->queue_lock);
543 		if (file)
544 			fput(file);
545 		return error;
546 	case NBD_SET_SOCK:
547 		if (lo->file)
548 			return -EBUSY;
549 		error = -EINVAL;
550 		file = fget(arg);
551 		if (file) {
552 			inode = file->f_dentry->d_inode;
553 			if (S_ISSOCK(inode->i_mode)) {
554 				lo->file = file;
555 				lo->sock = SOCKET_I(inode);
556 				error = 0;
557 			} else {
558 				fput(file);
559 			}
560 		}
561 		return error;
562 	case NBD_SET_BLKSIZE:
563 		lo->blksize = arg;
564 		lo->bytesize &= ~(lo->blksize-1);
565 		inode->i_bdev->bd_inode->i_size = lo->bytesize;
566 		set_blocksize(inode->i_bdev, lo->blksize);
567 		set_capacity(lo->disk, lo->bytesize >> 9);
568 		return 0;
569 	case NBD_SET_SIZE:
570 		lo->bytesize = arg & ~(lo->blksize-1);
571 		inode->i_bdev->bd_inode->i_size = lo->bytesize;
572 		set_blocksize(inode->i_bdev, lo->blksize);
573 		set_capacity(lo->disk, lo->bytesize >> 9);
574 		return 0;
575 	case NBD_SET_SIZE_BLOCKS:
576 		lo->bytesize = ((u64) arg) * lo->blksize;
577 		inode->i_bdev->bd_inode->i_size = lo->bytesize;
578 		set_blocksize(inode->i_bdev, lo->blksize);
579 		set_capacity(lo->disk, lo->bytesize >> 9);
580 		return 0;
581 	case NBD_DO_IT:
582 		if (!lo->file)
583 			return -EINVAL;
584 		nbd_do_it(lo);
585 		/* on return tidy up in case we have a signal */
586 		/* Forcibly shutdown the socket causing all listeners
587 		 * to error
588 		 *
589 		 * FIXME: This code is duplicated from sys_shutdown, but
590 		 * there should be a more generic interface rather than
591 		 * calling socket ops directly here */
592 		down(&lo->tx_lock);
593 		if (lo->sock) {
594 			printk(KERN_WARNING "%s: shutting down socket\n",
595 				lo->disk->disk_name);
596 			lo->sock->ops->shutdown(lo->sock,
597 				SEND_SHUTDOWN|RCV_SHUTDOWN);
598 			lo->sock = NULL;
599 		}
600 		up(&lo->tx_lock);
601 		spin_lock(&lo->queue_lock);
602 		file = lo->file;
603 		lo->file = NULL;
604 		spin_unlock(&lo->queue_lock);
605 		nbd_clear_que(lo);
606 		printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
607 		if (file)
608 			fput(file);
609 		return lo->harderror;
610 	case NBD_CLEAR_QUE:
611 		down(&lo->tx_lock);
612 		if (lo->sock) {
613 			up(&lo->tx_lock);
614 			return 0; /* probably should be error, but that would
615 				   * break "nbd-client -d", so just return 0 */
616 		}
617 		up(&lo->tx_lock);
618 		nbd_clear_que(lo);
619 		return 0;
620 	case NBD_PRINT_DEBUG:
621 		printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
622 			inode->i_bdev->bd_disk->disk_name,
623 			lo->queue_head.next, lo->queue_head.prev,
624 			&lo->queue_head);
625 		return 0;
626 	}
627 	return -EINVAL;
628 }
629 
630 static struct block_device_operations nbd_fops =
631 {
632 	.owner =	THIS_MODULE,
633 	.ioctl =	nbd_ioctl,
634 };
635 
636 /*
637  * And here should be modules and kernel interface
638  *  (Just smiley confuses emacs :-)
639  */
640 
641 static int __init nbd_init(void)
642 {
643 	int err = -ENOMEM;
644 	int i;
645 
646 	if (sizeof(struct nbd_request) != 28) {
647 		printk(KERN_CRIT "nbd: sizeof nbd_request needs to be 28 in order to work!\n" );
648 		return -EIO;
649 	}
650 
651 	if (nbds_max > MAX_NBD) {
652 		printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD,
653 				nbds_max);
654 		return -EINVAL;
655 	}
656 
657 	for (i = 0; i < nbds_max; i++) {
658 		struct gendisk *disk = alloc_disk(1);
659 		if (!disk)
660 			goto out;
661 		nbd_dev[i].disk = disk;
662 		/*
663 		 * The new linux 2.5 block layer implementation requires
664 		 * every gendisk to have its very own request_queue struct.
665 		 * These structs are big so we dynamically allocate them.
666 		 */
667 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
668 		if (!disk->queue) {
669 			put_disk(disk);
670 			goto out;
671 		}
672 	}
673 
674 	if (register_blkdev(NBD_MAJOR, "nbd")) {
675 		err = -EIO;
676 		goto out;
677 	}
678 
679 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
680 	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
681 
682 	devfs_mk_dir("nbd");
683 	for (i = 0; i < nbds_max; i++) {
684 		struct gendisk *disk = nbd_dev[i].disk;
685 		nbd_dev[i].file = NULL;
686 		nbd_dev[i].magic = LO_MAGIC;
687 		nbd_dev[i].flags = 0;
688 		spin_lock_init(&nbd_dev[i].queue_lock);
689 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
690 		init_MUTEX(&nbd_dev[i].tx_lock);
691 		nbd_dev[i].blksize = 1024;
692 		nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */
693 		disk->major = NBD_MAJOR;
694 		disk->first_minor = i;
695 		disk->fops = &nbd_fops;
696 		disk->private_data = &nbd_dev[i];
697 		disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
698 		sprintf(disk->disk_name, "nbd%d", i);
699 		sprintf(disk->devfs_name, "nbd/%d", i);
700 		set_capacity(disk, 0x7ffffc00ULL << 1); /* 2 TB */
701 		add_disk(disk);
702 	}
703 
704 	return 0;
705 out:
706 	while (i--) {
707 		blk_cleanup_queue(nbd_dev[i].disk->queue);
708 		put_disk(nbd_dev[i].disk);
709 	}
710 	return err;
711 }
712 
713 static void __exit nbd_cleanup(void)
714 {
715 	int i;
716 	for (i = 0; i < nbds_max; i++) {
717 		struct gendisk *disk = nbd_dev[i].disk;
718 		nbd_dev[i].magic = 0;
719 		if (disk) {
720 			del_gendisk(disk);
721 			blk_cleanup_queue(disk->queue);
722 			put_disk(disk);
723 		}
724 	}
725 	devfs_remove("nbd");
726 	unregister_blkdev(NBD_MAJOR, "nbd");
727 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
728 }
729 
730 module_init(nbd_init);
731 module_exit(nbd_cleanup);
732 
733 MODULE_DESCRIPTION("Network Block Device");
734 MODULE_LICENSE("GPL");
735 
736 module_param(nbds_max, int, 0444);
737 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
738 #ifndef NDEBUG
739 module_param(debugflags, int, 0644);
740 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
741 #endif
742