xref: /linux/drivers/block/nbd.c (revision f3d9478b2ce468c3115b02ecae7e975990697f15)
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 <linux/compiler.h>
58 #include <linux/err.h>
59 #include <linux/kernel.h>
60 #include <net/sock.h>
61 
62 #include <linux/devfs_fs_kernel.h>
63 
64 #include <asm/uaccess.h>
65 #include <asm/system.h>
66 #include <asm/types.h>
67 
68 #include <linux/nbd.h>
69 
70 #define LO_MAGIC 0x68797548
71 
72 #ifdef NDEBUG
73 #define dprintk(flags, fmt...)
74 #else /* NDEBUG */
75 #define dprintk(flags, fmt...) do { \
76 	if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
77 } while (0)
78 #define DBG_IOCTL       0x0004
79 #define DBG_INIT        0x0010
80 #define DBG_EXIT        0x0020
81 #define DBG_BLKDEV      0x0100
82 #define DBG_RX          0x0200
83 #define DBG_TX          0x0400
84 static unsigned int debugflags;
85 static unsigned int nbds_max = 16;
86 #endif /* NDEBUG */
87 
88 static struct nbd_device nbd_dev[MAX_NBD];
89 
90 /*
91  * Use just one lock (or at most 1 per NIC). Two arguments for this:
92  * 1. Each NIC is essentially a synchronization point for all servers
93  *    accessed through that NIC so there's no need to have more locks
94  *    than NICs anyway.
95  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
96  *    down each lock to the point where they're actually slower than just
97  *    a single lock.
98  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
99  */
100 static DEFINE_SPINLOCK(nbd_lock);
101 
102 #ifndef NDEBUG
103 static const char *ioctl_cmd_to_ascii(int cmd)
104 {
105 	switch (cmd) {
106 	case NBD_SET_SOCK: return "set-sock";
107 	case NBD_SET_BLKSIZE: return "set-blksize";
108 	case NBD_SET_SIZE: return "set-size";
109 	case NBD_DO_IT: return "do-it";
110 	case NBD_CLEAR_SOCK: return "clear-sock";
111 	case NBD_CLEAR_QUE: return "clear-que";
112 	case NBD_PRINT_DEBUG: return "print-debug";
113 	case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
114 	case NBD_DISCONNECT: return "disconnect";
115 	case BLKROSET: return "set-read-only";
116 	case BLKFLSBUF: return "flush-buffer-cache";
117 	}
118 	return "unknown";
119 }
120 
121 static const char *nbdcmd_to_ascii(int cmd)
122 {
123 	switch (cmd) {
124 	case  NBD_CMD_READ: return "read";
125 	case NBD_CMD_WRITE: return "write";
126 	case  NBD_CMD_DISC: return "disconnect";
127 	}
128 	return "invalid";
129 }
130 #endif /* NDEBUG */
131 
132 static void nbd_end_request(struct request *req)
133 {
134 	int uptodate = (req->errors == 0) ? 1 : 0;
135 	request_queue_t *q = req->q;
136 	unsigned long flags;
137 
138 	dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
139 			req, uptodate? "done": "failed");
140 
141 	spin_lock_irqsave(q->queue_lock, flags);
142 	if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
143 		end_that_request_last(req, uptodate);
144 	}
145 	spin_unlock_irqrestore(q->queue_lock, flags);
146 }
147 
148 /*
149  *  Send or receive packet.
150  */
151 static int sock_xmit(struct socket *sock, int send, void *buf, int size,
152 		int msg_flags)
153 {
154 	int result;
155 	struct msghdr msg;
156 	struct kvec iov;
157 	unsigned long flags;
158 	sigset_t oldset;
159 
160 	/* Allow interception of SIGKILL only
161 	 * Don't allow other signals to interrupt the transmission */
162 	spin_lock_irqsave(&current->sighand->siglock, flags);
163 	oldset = current->blocked;
164 	sigfillset(&current->blocked);
165 	sigdelsetmask(&current->blocked, sigmask(SIGKILL));
166 	recalc_sigpending();
167 	spin_unlock_irqrestore(&current->sighand->siglock, flags);
168 
169 	do {
170 		sock->sk->sk_allocation = GFP_NOIO;
171 		iov.iov_base = buf;
172 		iov.iov_len = size;
173 		msg.msg_name = NULL;
174 		msg.msg_namelen = 0;
175 		msg.msg_control = NULL;
176 		msg.msg_controllen = 0;
177 		msg.msg_flags = msg_flags | MSG_NOSIGNAL;
178 
179 		if (send)
180 			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
181 		else
182 			result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
183 
184 		if (signal_pending(current)) {
185 			siginfo_t info;
186 			spin_lock_irqsave(&current->sighand->siglock, flags);
187 			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
188 				current->pid, current->comm,
189 				dequeue_signal(current, &current->blocked, &info));
190 			spin_unlock_irqrestore(&current->sighand->siglock, flags);
191 			result = -EINTR;
192 			break;
193 		}
194 
195 		if (result <= 0) {
196 			if (result == 0)
197 				result = -EPIPE; /* short read */
198 			break;
199 		}
200 		size -= result;
201 		buf += result;
202 	} while (size > 0);
203 
204 	spin_lock_irqsave(&current->sighand->siglock, flags);
205 	current->blocked = oldset;
206 	recalc_sigpending();
207 	spin_unlock_irqrestore(&current->sighand->siglock, flags);
208 
209 	return result;
210 }
211 
212 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
213 		int flags)
214 {
215 	int result;
216 	void *kaddr = kmap(bvec->bv_page);
217 	result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len,
218 			flags);
219 	kunmap(bvec->bv_page);
220 	return result;
221 }
222 
223 static int nbd_send_req(struct nbd_device *lo, struct request *req)
224 {
225 	int result, i, flags;
226 	struct nbd_request request;
227 	unsigned long size = req->nr_sectors << 9;
228 	struct socket *sock = lo->sock;
229 
230 	request.magic = htonl(NBD_REQUEST_MAGIC);
231 	request.type = htonl(nbd_cmd(req));
232 	request.from = cpu_to_be64((u64) req->sector << 9);
233 	request.len = htonl(size);
234 	memcpy(request.handle, &req, sizeof(req));
235 
236 	dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
237 			lo->disk->disk_name, req,
238 			nbdcmd_to_ascii(nbd_cmd(req)),
239 			(unsigned long long)req->sector << 9,
240 			req->nr_sectors << 9);
241 	result = sock_xmit(sock, 1, &request, sizeof(request),
242 			(nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0);
243 	if (result <= 0) {
244 		printk(KERN_ERR "%s: Send control failed (result %d)\n",
245 				lo->disk->disk_name, result);
246 		goto error_out;
247 	}
248 
249 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
250 		struct bio *bio;
251 		/*
252 		 * we are really probing at internals to determine
253 		 * whether to set MSG_MORE or not...
254 		 */
255 		rq_for_each_bio(bio, req) {
256 			struct bio_vec *bvec;
257 			bio_for_each_segment(bvec, bio, i) {
258 				flags = 0;
259 				if ((i < (bio->bi_vcnt - 1)) || bio->bi_next)
260 					flags = MSG_MORE;
261 				dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
262 						lo->disk->disk_name, req,
263 						bvec->bv_len);
264 				result = sock_send_bvec(sock, bvec, flags);
265 				if (result <= 0) {
266 					printk(KERN_ERR "%s: Send data failed (result %d)\n",
267 							lo->disk->disk_name,
268 							result);
269 					goto error_out;
270 				}
271 			}
272 		}
273 	}
274 	return 0;
275 
276 error_out:
277 	return 1;
278 }
279 
280 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
281 {
282 	struct request *req;
283 	struct list_head *tmp;
284 	struct request *xreq;
285 	int err;
286 
287 	memcpy(&xreq, handle, sizeof(xreq));
288 
289 	err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
290 	if (unlikely(err))
291 		goto out;
292 
293 	spin_lock(&lo->queue_lock);
294 	list_for_each(tmp, &lo->queue_head) {
295 		req = list_entry(tmp, struct request, queuelist);
296 		if (req != xreq)
297 			continue;
298 		list_del_init(&req->queuelist);
299 		spin_unlock(&lo->queue_lock);
300 		return req;
301 	}
302 	spin_unlock(&lo->queue_lock);
303 
304 	err = -ENOENT;
305 
306 out:
307 	return ERR_PTR(err);
308 }
309 
310 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
311 {
312 	int result;
313 	void *kaddr = kmap(bvec->bv_page);
314 	result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len,
315 			MSG_WAITALL);
316 	kunmap(bvec->bv_page);
317 	return result;
318 }
319 
320 /* NULL returned = something went wrong, inform userspace */
321 static struct request *nbd_read_stat(struct nbd_device *lo)
322 {
323 	int result;
324 	struct nbd_reply reply;
325 	struct request *req;
326 	struct socket *sock = lo->sock;
327 
328 	reply.magic = 0;
329 	result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL);
330 	if (result <= 0) {
331 		printk(KERN_ERR "%s: Receive control failed (result %d)\n",
332 				lo->disk->disk_name, result);
333 		goto harderror;
334 	}
335 	req = nbd_find_request(lo, reply.handle);
336 	if (unlikely(IS_ERR(req))) {
337 		result = PTR_ERR(req);
338 		if (result != -ENOENT)
339 			goto harderror;
340 
341 		printk(KERN_ERR "%s: Unexpected reply (%p)\n",
342 				lo->disk->disk_name, reply.handle);
343 		result = -EBADR;
344 		goto harderror;
345 	}
346 
347 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
348 		printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
349 				lo->disk->disk_name,
350 				(unsigned long)ntohl(reply.magic));
351 		result = -EPROTO;
352 		goto harderror;
353 	}
354 	if (ntohl(reply.error)) {
355 		printk(KERN_ERR "%s: Other side returned error (%d)\n",
356 				lo->disk->disk_name, ntohl(reply.error));
357 		req->errors++;
358 		return req;
359 	}
360 
361 	dprintk(DBG_RX, "%s: request %p: got reply\n",
362 			lo->disk->disk_name, req);
363 	if (nbd_cmd(req) == NBD_CMD_READ) {
364 		int i;
365 		struct bio *bio;
366 		rq_for_each_bio(bio, req) {
367 			struct bio_vec *bvec;
368 			bio_for_each_segment(bvec, bio, i) {
369 				result = sock_recv_bvec(sock, bvec);
370 				if (result <= 0) {
371 					printk(KERN_ERR "%s: Receive data failed (result %d)\n",
372 							lo->disk->disk_name,
373 							result);
374 					goto harderror;
375 				}
376 				dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
377 					lo->disk->disk_name, req, bvec->bv_len);
378 			}
379 		}
380 	}
381 	return req;
382 harderror:
383 	lo->harderror = result;
384 	return NULL;
385 }
386 
387 static void nbd_do_it(struct nbd_device *lo)
388 {
389 	struct request *req;
390 
391 	BUG_ON(lo->magic != LO_MAGIC);
392 
393 	while ((req = nbd_read_stat(lo)) != NULL)
394 		nbd_end_request(req);
395 	return;
396 }
397 
398 static void nbd_clear_que(struct nbd_device *lo)
399 {
400 	struct request *req;
401 
402 	BUG_ON(lo->magic != LO_MAGIC);
403 
404 	/*
405 	 * Because we have set lo->sock to NULL under the tx_lock, all
406 	 * modifications to the list must have completed by now.  For
407 	 * the same reason, the active_req must be NULL.
408 	 *
409 	 * As a consequence, we don't need to take the spin lock while
410 	 * purging the list here.
411 	 */
412 	BUG_ON(lo->sock);
413 	BUG_ON(lo->active_req);
414 
415 	while (!list_empty(&lo->queue_head)) {
416 		req = list_entry(lo->queue_head.next, struct request,
417 				 queuelist);
418 		list_del_init(&req->queuelist);
419 		req->errors++;
420 		nbd_end_request(req);
421 	}
422 }
423 
424 /*
425  * We always wait for result of write, for now. It would be nice to make it optional
426  * in future
427  * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
428  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
429  */
430 
431 static void do_nbd_request(request_queue_t * q)
432 {
433 	struct request *req;
434 
435 	while ((req = elv_next_request(q)) != NULL) {
436 		struct nbd_device *lo;
437 
438 		blkdev_dequeue_request(req);
439 		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%lx)\n",
440 				req->rq_disk->disk_name, req, req->flags);
441 
442 		if (!(req->flags & REQ_CMD))
443 			goto error_out;
444 
445 		lo = req->rq_disk->private_data;
446 
447 		BUG_ON(lo->magic != LO_MAGIC);
448 
449 		nbd_cmd(req) = NBD_CMD_READ;
450 		if (rq_data_dir(req) == WRITE) {
451 			nbd_cmd(req) = NBD_CMD_WRITE;
452 			if (lo->flags & NBD_READ_ONLY) {
453 				printk(KERN_ERR "%s: Write on read-only\n",
454 						lo->disk->disk_name);
455 				goto error_out;
456 			}
457 		}
458 
459 		req->errors = 0;
460 		spin_unlock_irq(q->queue_lock);
461 
462 		mutex_lock(&lo->tx_lock);
463 		if (unlikely(!lo->sock)) {
464 			mutex_unlock(&lo->tx_lock);
465 			printk(KERN_ERR "%s: Attempted send on closed socket\n",
466 			       lo->disk->disk_name);
467 			req->errors++;
468 			nbd_end_request(req);
469 			spin_lock_irq(q->queue_lock);
470 			continue;
471 		}
472 
473 		lo->active_req = req;
474 
475 		if (nbd_send_req(lo, req) != 0) {
476 			printk(KERN_ERR "%s: Request send failed\n",
477 					lo->disk->disk_name);
478 			req->errors++;
479 			nbd_end_request(req);
480 		} else {
481 			spin_lock(&lo->queue_lock);
482 			list_add(&req->queuelist, &lo->queue_head);
483 			spin_unlock(&lo->queue_lock);
484 		}
485 
486 		lo->active_req = NULL;
487 		mutex_unlock(&lo->tx_lock);
488 		wake_up_all(&lo->active_wq);
489 
490 		spin_lock_irq(q->queue_lock);
491 		continue;
492 
493 error_out:
494 		req->errors++;
495 		spin_unlock(q->queue_lock);
496 		nbd_end_request(req);
497 		spin_lock(q->queue_lock);
498 	}
499 	return;
500 }
501 
502 static int nbd_ioctl(struct inode *inode, struct file *file,
503 		     unsigned int cmd, unsigned long arg)
504 {
505 	struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
506 	int error;
507 	struct request sreq ;
508 
509 	if (!capable(CAP_SYS_ADMIN))
510 		return -EPERM;
511 
512 	BUG_ON(lo->magic != LO_MAGIC);
513 
514 	/* Anyone capable of this syscall can do *real bad* things */
515 	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
516 			lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
517 
518 	switch (cmd) {
519 	case NBD_DISCONNECT:
520 	        printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
521 		sreq.flags = REQ_SPECIAL;
522 		nbd_cmd(&sreq) = NBD_CMD_DISC;
523 		/*
524 		 * Set these to sane values in case server implementation
525 		 * fails to check the request type first and also to keep
526 		 * debugging output cleaner.
527 		 */
528 		sreq.sector = 0;
529 		sreq.nr_sectors = 0;
530                 if (!lo->sock)
531 			return -EINVAL;
532                 nbd_send_req(lo, &sreq);
533                 return 0;
534 
535 	case NBD_CLEAR_SOCK:
536 		error = 0;
537 		mutex_lock(&lo->tx_lock);
538 		lo->sock = NULL;
539 		mutex_unlock(&lo->tx_lock);
540 		file = lo->file;
541 		lo->file = NULL;
542 		nbd_clear_que(lo);
543 		BUG_ON(!list_empty(&lo->queue_head));
544 		if (file)
545 			fput(file);
546 		return error;
547 	case NBD_SET_SOCK:
548 		if (lo->file)
549 			return -EBUSY;
550 		error = -EINVAL;
551 		file = fget(arg);
552 		if (file) {
553 			inode = file->f_dentry->d_inode;
554 			if (S_ISSOCK(inode->i_mode)) {
555 				lo->file = file;
556 				lo->sock = SOCKET_I(inode);
557 				error = 0;
558 			} else {
559 				fput(file);
560 			}
561 		}
562 		return error;
563 	case NBD_SET_BLKSIZE:
564 		lo->blksize = arg;
565 		lo->bytesize &= ~(lo->blksize-1);
566 		inode->i_bdev->bd_inode->i_size = lo->bytesize;
567 		set_blocksize(inode->i_bdev, lo->blksize);
568 		set_capacity(lo->disk, lo->bytesize >> 9);
569 		return 0;
570 	case NBD_SET_SIZE:
571 		lo->bytesize = arg & ~(lo->blksize-1);
572 		inode->i_bdev->bd_inode->i_size = lo->bytesize;
573 		set_blocksize(inode->i_bdev, lo->blksize);
574 		set_capacity(lo->disk, lo->bytesize >> 9);
575 		return 0;
576 	case NBD_SET_SIZE_BLOCKS:
577 		lo->bytesize = ((u64) arg) * lo->blksize;
578 		inode->i_bdev->bd_inode->i_size = lo->bytesize;
579 		set_blocksize(inode->i_bdev, lo->blksize);
580 		set_capacity(lo->disk, lo->bytesize >> 9);
581 		return 0;
582 	case NBD_DO_IT:
583 		if (!lo->file)
584 			return -EINVAL;
585 		nbd_do_it(lo);
586 		/* on return tidy up in case we have a signal */
587 		/* Forcibly shutdown the socket causing all listeners
588 		 * to error
589 		 *
590 		 * FIXME: This code is duplicated from sys_shutdown, but
591 		 * there should be a more generic interface rather than
592 		 * calling socket ops directly here */
593 		mutex_lock(&lo->tx_lock);
594 		if (lo->sock) {
595 			printk(KERN_WARNING "%s: shutting down socket\n",
596 				lo->disk->disk_name);
597 			lo->sock->ops->shutdown(lo->sock,
598 				SEND_SHUTDOWN|RCV_SHUTDOWN);
599 			lo->sock = NULL;
600 		}
601 		mutex_unlock(&lo->tx_lock);
602 		file = lo->file;
603 		lo->file = NULL;
604 		nbd_clear_que(lo);
605 		printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
606 		if (file)
607 			fput(file);
608 		return lo->harderror;
609 	case NBD_CLEAR_QUE:
610 		/*
611 		 * This is for compatibility only.  The queue is always cleared
612 		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
613 		 */
614 		BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
615 		return 0;
616 	case NBD_PRINT_DEBUG:
617 		printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
618 			inode->i_bdev->bd_disk->disk_name,
619 			lo->queue_head.next, lo->queue_head.prev,
620 			&lo->queue_head);
621 		return 0;
622 	}
623 	return -EINVAL;
624 }
625 
626 static struct block_device_operations nbd_fops =
627 {
628 	.owner =	THIS_MODULE,
629 	.ioctl =	nbd_ioctl,
630 };
631 
632 /*
633  * And here should be modules and kernel interface
634  *  (Just smiley confuses emacs :-)
635  */
636 
637 static int __init nbd_init(void)
638 {
639 	int err = -ENOMEM;
640 	int i;
641 
642 	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
643 
644 	if (nbds_max > MAX_NBD) {
645 		printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD,
646 				nbds_max);
647 		return -EINVAL;
648 	}
649 
650 	for (i = 0; i < nbds_max; i++) {
651 		struct gendisk *disk = alloc_disk(1);
652 		if (!disk)
653 			goto out;
654 		nbd_dev[i].disk = disk;
655 		/*
656 		 * The new linux 2.5 block layer implementation requires
657 		 * every gendisk to have its very own request_queue struct.
658 		 * These structs are big so we dynamically allocate them.
659 		 */
660 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
661 		if (!disk->queue) {
662 			put_disk(disk);
663 			goto out;
664 		}
665 	}
666 
667 	if (register_blkdev(NBD_MAJOR, "nbd")) {
668 		err = -EIO;
669 		goto out;
670 	}
671 
672 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
673 	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
674 
675 	devfs_mk_dir("nbd");
676 	for (i = 0; i < nbds_max; i++) {
677 		struct gendisk *disk = nbd_dev[i].disk;
678 		nbd_dev[i].file = NULL;
679 		nbd_dev[i].magic = LO_MAGIC;
680 		nbd_dev[i].flags = 0;
681 		spin_lock_init(&nbd_dev[i].queue_lock);
682 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
683 		mutex_init(&nbd_dev[i].tx_lock);
684 		init_waitqueue_head(&nbd_dev[i].active_wq);
685 		nbd_dev[i].blksize = 1024;
686 		nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */
687 		disk->major = NBD_MAJOR;
688 		disk->first_minor = i;
689 		disk->fops = &nbd_fops;
690 		disk->private_data = &nbd_dev[i];
691 		disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
692 		sprintf(disk->disk_name, "nbd%d", i);
693 		sprintf(disk->devfs_name, "nbd/%d", i);
694 		set_capacity(disk, 0x7ffffc00ULL << 1); /* 2 TB */
695 		add_disk(disk);
696 	}
697 
698 	return 0;
699 out:
700 	while (i--) {
701 		blk_cleanup_queue(nbd_dev[i].disk->queue);
702 		put_disk(nbd_dev[i].disk);
703 	}
704 	return err;
705 }
706 
707 static void __exit nbd_cleanup(void)
708 {
709 	int i;
710 	for (i = 0; i < nbds_max; i++) {
711 		struct gendisk *disk = nbd_dev[i].disk;
712 		nbd_dev[i].magic = 0;
713 		if (disk) {
714 			del_gendisk(disk);
715 			blk_cleanup_queue(disk->queue);
716 			put_disk(disk);
717 		}
718 	}
719 	devfs_remove("nbd");
720 	unregister_blkdev(NBD_MAJOR, "nbd");
721 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
722 }
723 
724 module_init(nbd_init);
725 module_exit(nbd_cleanup);
726 
727 MODULE_DESCRIPTION("Network Block Device");
728 MODULE_LICENSE("GPL");
729 
730 module_param(nbds_max, int, 0444);
731 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
732 #ifndef NDEBUG
733 module_param(debugflags, int, 0644);
734 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
735 #endif
736