xref: /openbmc/linux/drivers/block/nbd.c (revision b8bb76713ec50df2f11efee386e16f93d51e1076)
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 #include <linux/net.h>
32 #include <linux/kthread.h>
33 
34 #include <asm/uaccess.h>
35 #include <asm/system.h>
36 #include <asm/types.h>
37 
38 #include <linux/nbd.h>
39 
40 #define LO_MAGIC 0x68797548
41 
42 #ifdef NDEBUG
43 #define dprintk(flags, fmt...)
44 #else /* NDEBUG */
45 #define dprintk(flags, fmt...) do { \
46 	if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
47 } while (0)
48 #define DBG_IOCTL       0x0004
49 #define DBG_INIT        0x0010
50 #define DBG_EXIT        0x0020
51 #define DBG_BLKDEV      0x0100
52 #define DBG_RX          0x0200
53 #define DBG_TX          0x0400
54 static unsigned int debugflags;
55 #endif /* NDEBUG */
56 
57 static unsigned int nbds_max = 16;
58 static struct nbd_device *nbd_dev;
59 static int max_part;
60 
61 /*
62  * Use just one lock (or at most 1 per NIC). Two arguments for this:
63  * 1. Each NIC is essentially a synchronization point for all servers
64  *    accessed through that NIC so there's no need to have more locks
65  *    than NICs anyway.
66  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
67  *    down each lock to the point where they're actually slower than just
68  *    a single lock.
69  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
70  */
71 static DEFINE_SPINLOCK(nbd_lock);
72 
73 #ifndef NDEBUG
74 static const char *ioctl_cmd_to_ascii(int cmd)
75 {
76 	switch (cmd) {
77 	case NBD_SET_SOCK: return "set-sock";
78 	case NBD_SET_BLKSIZE: return "set-blksize";
79 	case NBD_SET_SIZE: return "set-size";
80 	case NBD_DO_IT: return "do-it";
81 	case NBD_CLEAR_SOCK: return "clear-sock";
82 	case NBD_CLEAR_QUE: return "clear-que";
83 	case NBD_PRINT_DEBUG: return "print-debug";
84 	case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
85 	case NBD_DISCONNECT: return "disconnect";
86 	case BLKROSET: return "set-read-only";
87 	case BLKFLSBUF: return "flush-buffer-cache";
88 	}
89 	return "unknown";
90 }
91 
92 static const char *nbdcmd_to_ascii(int cmd)
93 {
94 	switch (cmd) {
95 	case  NBD_CMD_READ: return "read";
96 	case NBD_CMD_WRITE: return "write";
97 	case  NBD_CMD_DISC: return "disconnect";
98 	}
99 	return "invalid";
100 }
101 #endif /* NDEBUG */
102 
103 static void nbd_end_request(struct request *req)
104 {
105 	int error = req->errors ? -EIO : 0;
106 	struct request_queue *q = req->q;
107 	unsigned long flags;
108 
109 	dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
110 			req, error ? "failed" : "done");
111 
112 	spin_lock_irqsave(q->queue_lock, flags);
113 	__blk_end_request(req, error, req->nr_sectors << 9);
114 	spin_unlock_irqrestore(q->queue_lock, flags);
115 }
116 
117 static void sock_shutdown(struct nbd_device *lo, int lock)
118 {
119 	/* Forcibly shutdown the socket causing all listeners
120 	 * to error
121 	 *
122 	 * FIXME: This code is duplicated from sys_shutdown, but
123 	 * there should be a more generic interface rather than
124 	 * calling socket ops directly here */
125 	if (lock)
126 		mutex_lock(&lo->tx_lock);
127 	if (lo->sock) {
128 		printk(KERN_WARNING "%s: shutting down socket\n",
129 			lo->disk->disk_name);
130 		kernel_sock_shutdown(lo->sock, SHUT_RDWR);
131 		lo->sock = NULL;
132 	}
133 	if (lock)
134 		mutex_unlock(&lo->tx_lock);
135 }
136 
137 static void nbd_xmit_timeout(unsigned long arg)
138 {
139 	struct task_struct *task = (struct task_struct *)arg;
140 
141 	printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
142 		task->comm, task->pid);
143 	force_sig(SIGKILL, task);
144 }
145 
146 /*
147  *  Send or receive packet.
148  */
149 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
150 		int msg_flags)
151 {
152 	struct socket *sock = lo->sock;
153 	int result;
154 	struct msghdr msg;
155 	struct kvec iov;
156 	sigset_t blocked, oldset;
157 
158 	if (unlikely(!sock)) {
159 		printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
160 		       lo->disk->disk_name, (send ? "send" : "recv"));
161 		return -EINVAL;
162 	}
163 
164 	/* Allow interception of SIGKILL only
165 	 * Don't allow other signals to interrupt the transmission */
166 	siginitsetinv(&blocked, sigmask(SIGKILL));
167 	sigprocmask(SIG_SETMASK, &blocked, &oldset);
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 			struct timer_list ti;
181 
182 			if (lo->xmit_timeout) {
183 				init_timer(&ti);
184 				ti.function = nbd_xmit_timeout;
185 				ti.data = (unsigned long)current;
186 				ti.expires = jiffies + lo->xmit_timeout;
187 				add_timer(&ti);
188 			}
189 			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
190 			if (lo->xmit_timeout)
191 				del_timer_sync(&ti);
192 		} else
193 			result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
194 
195 		if (signal_pending(current)) {
196 			siginfo_t info;
197 			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
198 				task_pid_nr(current), current->comm,
199 				dequeue_signal_lock(current, &current->blocked, &info));
200 			result = -EINTR;
201 			sock_shutdown(lo, !send);
202 			break;
203 		}
204 
205 		if (result <= 0) {
206 			if (result == 0)
207 				result = -EPIPE; /* short read */
208 			break;
209 		}
210 		size -= result;
211 		buf += result;
212 	} while (size > 0);
213 
214 	sigprocmask(SIG_SETMASK, &oldset, NULL);
215 
216 	return result;
217 }
218 
219 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
220 		int flags)
221 {
222 	int result;
223 	void *kaddr = kmap(bvec->bv_page);
224 	result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
225 	kunmap(bvec->bv_page);
226 	return result;
227 }
228 
229 /* always call with the tx_lock held */
230 static int nbd_send_req(struct nbd_device *lo, struct request *req)
231 {
232 	int result, flags;
233 	struct nbd_request request;
234 	unsigned long size = req->nr_sectors << 9;
235 
236 	request.magic = htonl(NBD_REQUEST_MAGIC);
237 	request.type = htonl(nbd_cmd(req));
238 	request.from = cpu_to_be64((u64) req->sector << 9);
239 	request.len = htonl(size);
240 	memcpy(request.handle, &req, sizeof(req));
241 
242 	dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
243 			lo->disk->disk_name, req,
244 			nbdcmd_to_ascii(nbd_cmd(req)),
245 			(unsigned long long)req->sector << 9,
246 			req->nr_sectors << 9);
247 	result = sock_xmit(lo, 1, &request, sizeof(request),
248 			(nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
249 	if (result <= 0) {
250 		printk(KERN_ERR "%s: Send control failed (result %d)\n",
251 				lo->disk->disk_name, result);
252 		goto error_out;
253 	}
254 
255 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
256 		struct req_iterator iter;
257 		struct bio_vec *bvec;
258 		/*
259 		 * we are really probing at internals to determine
260 		 * whether to set MSG_MORE or not...
261 		 */
262 		rq_for_each_segment(bvec, req, iter) {
263 			flags = 0;
264 			if (!rq_iter_last(req, iter))
265 				flags = MSG_MORE;
266 			dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
267 					lo->disk->disk_name, req, bvec->bv_len);
268 			result = sock_send_bvec(lo, bvec, flags);
269 			if (result <= 0) {
270 				printk(KERN_ERR "%s: Send data failed (result %d)\n",
271 						lo->disk->disk_name, result);
272 				goto error_out;
273 			}
274 		}
275 	}
276 	return 0;
277 
278 error_out:
279 	return 1;
280 }
281 
282 static struct request *nbd_find_request(struct nbd_device *lo,
283 					struct request *xreq)
284 {
285 	struct request *req, *tmp;
286 	int err;
287 
288 	err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
289 	if (unlikely(err))
290 		goto out;
291 
292 	spin_lock(&lo->queue_lock);
293 	list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
294 		if (req != xreq)
295 			continue;
296 		list_del_init(&req->queuelist);
297 		spin_unlock(&lo->queue_lock);
298 		return req;
299 	}
300 	spin_unlock(&lo->queue_lock);
301 
302 	err = -ENOENT;
303 
304 out:
305 	return ERR_PTR(err);
306 }
307 
308 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
309 {
310 	int result;
311 	void *kaddr = kmap(bvec->bv_page);
312 	result = sock_xmit(lo, 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 
325 	reply.magic = 0;
326 	result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
327 	if (result <= 0) {
328 		printk(KERN_ERR "%s: Receive control failed (result %d)\n",
329 				lo->disk->disk_name, result);
330 		goto harderror;
331 	}
332 
333 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
334 		printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
335 				lo->disk->disk_name,
336 				(unsigned long)ntohl(reply.magic));
337 		result = -EPROTO;
338 		goto harderror;
339 	}
340 
341 	req = nbd_find_request(lo, *(struct request **)reply.handle);
342 	if (IS_ERR(req)) {
343 		result = PTR_ERR(req);
344 		if (result != -ENOENT)
345 			goto harderror;
346 
347 		printk(KERN_ERR "%s: Unexpected reply (%p)\n",
348 				lo->disk->disk_name, reply.handle);
349 		result = -EBADR;
350 		goto harderror;
351 	}
352 
353 	if (ntohl(reply.error)) {
354 		printk(KERN_ERR "%s: Other side returned error (%d)\n",
355 				lo->disk->disk_name, ntohl(reply.error));
356 		req->errors++;
357 		return req;
358 	}
359 
360 	dprintk(DBG_RX, "%s: request %p: got reply\n",
361 			lo->disk->disk_name, req);
362 	if (nbd_cmd(req) == NBD_CMD_READ) {
363 		struct req_iterator iter;
364 		struct bio_vec *bvec;
365 
366 		rq_for_each_segment(bvec, req, iter) {
367 			result = sock_recv_bvec(lo, bvec);
368 			if (result <= 0) {
369 				printk(KERN_ERR "%s: Receive data failed (result %d)\n",
370 						lo->disk->disk_name, result);
371 				req->errors++;
372 				return req;
373 			}
374 			dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
375 				lo->disk->disk_name, req, bvec->bv_len);
376 		}
377 	}
378 	return req;
379 harderror:
380 	lo->harderror = result;
381 	return NULL;
382 }
383 
384 static ssize_t pid_show(struct device *dev,
385 			struct device_attribute *attr, char *buf)
386 {
387 	struct gendisk *disk = dev_to_disk(dev);
388 
389 	return sprintf(buf, "%ld\n",
390 		(long) ((struct nbd_device *)disk->private_data)->pid);
391 }
392 
393 static struct device_attribute pid_attr = {
394 	.attr = { .name = "pid", .mode = S_IRUGO},
395 	.show = pid_show,
396 };
397 
398 static int nbd_do_it(struct nbd_device *lo)
399 {
400 	struct request *req;
401 	int ret;
402 
403 	BUG_ON(lo->magic != LO_MAGIC);
404 
405 	lo->pid = current->pid;
406 	ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
407 	if (ret) {
408 		printk(KERN_ERR "nbd: sysfs_create_file failed!");
409 		lo->pid = 0;
410 		return ret;
411 	}
412 
413 	while ((req = nbd_read_stat(lo)) != NULL)
414 		nbd_end_request(req);
415 
416 	sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
417 	lo->pid = 0;
418 	return 0;
419 }
420 
421 static void nbd_clear_que(struct nbd_device *lo)
422 {
423 	struct request *req;
424 
425 	BUG_ON(lo->magic != LO_MAGIC);
426 
427 	/*
428 	 * Because we have set lo->sock to NULL under the tx_lock, all
429 	 * modifications to the list must have completed by now.  For
430 	 * the same reason, the active_req must be NULL.
431 	 *
432 	 * As a consequence, we don't need to take the spin lock while
433 	 * purging the list here.
434 	 */
435 	BUG_ON(lo->sock);
436 	BUG_ON(lo->active_req);
437 
438 	while (!list_empty(&lo->queue_head)) {
439 		req = list_entry(lo->queue_head.next, struct request,
440 				 queuelist);
441 		list_del_init(&req->queuelist);
442 		req->errors++;
443 		nbd_end_request(req);
444 	}
445 }
446 
447 
448 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
449 {
450 	if (!blk_fs_request(req))
451 		goto error_out;
452 
453 	nbd_cmd(req) = NBD_CMD_READ;
454 	if (rq_data_dir(req) == WRITE) {
455 		nbd_cmd(req) = NBD_CMD_WRITE;
456 		if (lo->flags & NBD_READ_ONLY) {
457 			printk(KERN_ERR "%s: Write on read-only\n",
458 					lo->disk->disk_name);
459 			goto error_out;
460 		}
461 	}
462 
463 	req->errors = 0;
464 
465 	mutex_lock(&lo->tx_lock);
466 	if (unlikely(!lo->sock)) {
467 		mutex_unlock(&lo->tx_lock);
468 		printk(KERN_ERR "%s: Attempted send on closed socket\n",
469 		       lo->disk->disk_name);
470 		req->errors++;
471 		nbd_end_request(req);
472 		return;
473 	}
474 
475 	lo->active_req = req;
476 
477 	if (nbd_send_req(lo, req) != 0) {
478 		printk(KERN_ERR "%s: Request send failed\n",
479 				lo->disk->disk_name);
480 		req->errors++;
481 		nbd_end_request(req);
482 	} else {
483 		spin_lock(&lo->queue_lock);
484 		list_add(&req->queuelist, &lo->queue_head);
485 		spin_unlock(&lo->queue_lock);
486 	}
487 
488 	lo->active_req = NULL;
489 	mutex_unlock(&lo->tx_lock);
490 	wake_up_all(&lo->active_wq);
491 
492 	return;
493 
494 error_out:
495 	req->errors++;
496 	nbd_end_request(req);
497 }
498 
499 static int nbd_thread(void *data)
500 {
501 	struct nbd_device *lo = data;
502 	struct request *req;
503 
504 	set_user_nice(current, -20);
505 	while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
506 		/* wait for something to do */
507 		wait_event_interruptible(lo->waiting_wq,
508 					 kthread_should_stop() ||
509 					 !list_empty(&lo->waiting_queue));
510 
511 		/* extract request */
512 		if (list_empty(&lo->waiting_queue))
513 			continue;
514 
515 		spin_lock_irq(&lo->queue_lock);
516 		req = list_entry(lo->waiting_queue.next, struct request,
517 				 queuelist);
518 		list_del_init(&req->queuelist);
519 		spin_unlock_irq(&lo->queue_lock);
520 
521 		/* handle request */
522 		nbd_handle_req(lo, req);
523 	}
524 	return 0;
525 }
526 
527 /*
528  * We always wait for result of write, for now. It would be nice to make it optional
529  * in future
530  * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
531  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
532  */
533 
534 static void do_nbd_request(struct request_queue * q)
535 {
536 	struct request *req;
537 
538 	while ((req = elv_next_request(q)) != NULL) {
539 		struct nbd_device *lo;
540 
541 		blkdev_dequeue_request(req);
542 
543 		spin_unlock_irq(q->queue_lock);
544 
545 		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
546 				req->rq_disk->disk_name, req, req->cmd_type);
547 
548 		lo = req->rq_disk->private_data;
549 
550 		BUG_ON(lo->magic != LO_MAGIC);
551 
552 		if (unlikely(!lo->sock)) {
553 			printk(KERN_ERR "%s: Attempted send on closed socket\n",
554 				lo->disk->disk_name);
555 			req->errors++;
556 			nbd_end_request(req);
557 			spin_lock_irq(q->queue_lock);
558 			continue;
559 		}
560 
561 		spin_lock_irq(&lo->queue_lock);
562 		list_add_tail(&req->queuelist, &lo->waiting_queue);
563 		spin_unlock_irq(&lo->queue_lock);
564 
565 		wake_up(&lo->waiting_wq);
566 
567 		spin_lock_irq(q->queue_lock);
568 	}
569 }
570 
571 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
572 		     unsigned int cmd, unsigned long arg)
573 {
574 	struct nbd_device *lo = bdev->bd_disk->private_data;
575 	struct file *file;
576 	int error;
577 	struct request sreq ;
578 	struct task_struct *thread;
579 
580 	if (!capable(CAP_SYS_ADMIN))
581 		return -EPERM;
582 
583 	BUG_ON(lo->magic != LO_MAGIC);
584 
585 	/* Anyone capable of this syscall can do *real bad* things */
586 	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
587 			lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
588 
589 	switch (cmd) {
590 	case NBD_DISCONNECT:
591 	        printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
592 		blk_rq_init(NULL, &sreq);
593 		sreq.cmd_type = REQ_TYPE_SPECIAL;
594 		nbd_cmd(&sreq) = NBD_CMD_DISC;
595 		/*
596 		 * Set these to sane values in case server implementation
597 		 * fails to check the request type first and also to keep
598 		 * debugging output cleaner.
599 		 */
600 		sreq.sector = 0;
601 		sreq.nr_sectors = 0;
602                 if (!lo->sock)
603 			return -EINVAL;
604 		mutex_lock(&lo->tx_lock);
605                 nbd_send_req(lo, &sreq);
606 		mutex_unlock(&lo->tx_lock);
607                 return 0;
608 
609 	case NBD_CLEAR_SOCK:
610 		error = 0;
611 		mutex_lock(&lo->tx_lock);
612 		lo->sock = NULL;
613 		mutex_unlock(&lo->tx_lock);
614 		file = lo->file;
615 		lo->file = NULL;
616 		nbd_clear_que(lo);
617 		BUG_ON(!list_empty(&lo->queue_head));
618 		if (file)
619 			fput(file);
620 		return error;
621 	case NBD_SET_SOCK:
622 		if (lo->file)
623 			return -EBUSY;
624 		error = -EINVAL;
625 		file = fget(arg);
626 		if (file) {
627 			struct inode *inode = file->f_path.dentry->d_inode;
628 			if (S_ISSOCK(inode->i_mode)) {
629 				lo->file = file;
630 				lo->sock = SOCKET_I(inode);
631 				if (max_part > 0)
632 					bdev->bd_invalidated = 1;
633 				error = 0;
634 			} else {
635 				fput(file);
636 			}
637 		}
638 		return error;
639 	case NBD_SET_BLKSIZE:
640 		lo->blksize = arg;
641 		lo->bytesize &= ~(lo->blksize-1);
642 		bdev->bd_inode->i_size = lo->bytesize;
643 		set_blocksize(bdev, lo->blksize);
644 		set_capacity(lo->disk, lo->bytesize >> 9);
645 		return 0;
646 	case NBD_SET_SIZE:
647 		lo->bytesize = arg & ~(lo->blksize-1);
648 		bdev->bd_inode->i_size = lo->bytesize;
649 		set_blocksize(bdev, lo->blksize);
650 		set_capacity(lo->disk, lo->bytesize >> 9);
651 		return 0;
652 	case NBD_SET_TIMEOUT:
653 		lo->xmit_timeout = arg * HZ;
654 		return 0;
655 	case NBD_SET_SIZE_BLOCKS:
656 		lo->bytesize = ((u64) arg) * lo->blksize;
657 		bdev->bd_inode->i_size = lo->bytesize;
658 		set_blocksize(bdev, lo->blksize);
659 		set_capacity(lo->disk, lo->bytesize >> 9);
660 		return 0;
661 	case NBD_DO_IT:
662 		if (lo->pid)
663 			return -EBUSY;
664 		if (!lo->file)
665 			return -EINVAL;
666 		thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
667 		if (IS_ERR(thread))
668 			return PTR_ERR(thread);
669 		wake_up_process(thread);
670 		error = nbd_do_it(lo);
671 		kthread_stop(thread);
672 		if (error)
673 			return error;
674 		sock_shutdown(lo, 1);
675 		file = lo->file;
676 		lo->file = NULL;
677 		nbd_clear_que(lo);
678 		printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
679 		if (file)
680 			fput(file);
681 		lo->bytesize = 0;
682 		bdev->bd_inode->i_size = 0;
683 		set_capacity(lo->disk, 0);
684 		if (max_part > 0)
685 			ioctl_by_bdev(bdev, BLKRRPART, 0);
686 		return lo->harderror;
687 	case NBD_CLEAR_QUE:
688 		/*
689 		 * This is for compatibility only.  The queue is always cleared
690 		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
691 		 */
692 		BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
693 		return 0;
694 	case NBD_PRINT_DEBUG:
695 		printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
696 			bdev->bd_disk->disk_name,
697 			lo->queue_head.next, lo->queue_head.prev,
698 			&lo->queue_head);
699 		return 0;
700 	}
701 	return -EINVAL;
702 }
703 
704 static struct block_device_operations nbd_fops =
705 {
706 	.owner =	THIS_MODULE,
707 	.locked_ioctl =	nbd_ioctl,
708 };
709 
710 /*
711  * And here should be modules and kernel interface
712  *  (Just smiley confuses emacs :-)
713  */
714 
715 static int __init nbd_init(void)
716 {
717 	int err = -ENOMEM;
718 	int i;
719 	int part_shift;
720 
721 	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
722 
723 	if (max_part < 0) {
724 		printk(KERN_CRIT "nbd: max_part must be >= 0\n");
725 		return -EINVAL;
726 	}
727 
728 	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
729 	if (!nbd_dev)
730 		return -ENOMEM;
731 
732 	part_shift = 0;
733 	if (max_part > 0)
734 		part_shift = fls(max_part);
735 
736 	for (i = 0; i < nbds_max; i++) {
737 		struct gendisk *disk = alloc_disk(1 << part_shift);
738 		if (!disk)
739 			goto out;
740 		nbd_dev[i].disk = disk;
741 		/*
742 		 * The new linux 2.5 block layer implementation requires
743 		 * every gendisk to have its very own request_queue struct.
744 		 * These structs are big so we dynamically allocate them.
745 		 */
746 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
747 		if (!disk->queue) {
748 			put_disk(disk);
749 			goto out;
750 		}
751 		/*
752 		 * Tell the block layer that we are not a rotational device
753 		 */
754 		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
755 	}
756 
757 	if (register_blkdev(NBD_MAJOR, "nbd")) {
758 		err = -EIO;
759 		goto out;
760 	}
761 
762 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
763 	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
764 
765 	for (i = 0; i < nbds_max; i++) {
766 		struct gendisk *disk = nbd_dev[i].disk;
767 		nbd_dev[i].file = NULL;
768 		nbd_dev[i].magic = LO_MAGIC;
769 		nbd_dev[i].flags = 0;
770 		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
771 		spin_lock_init(&nbd_dev[i].queue_lock);
772 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
773 		mutex_init(&nbd_dev[i].tx_lock);
774 		init_waitqueue_head(&nbd_dev[i].active_wq);
775 		init_waitqueue_head(&nbd_dev[i].waiting_wq);
776 		nbd_dev[i].blksize = 1024;
777 		nbd_dev[i].bytesize = 0;
778 		disk->major = NBD_MAJOR;
779 		disk->first_minor = i << part_shift;
780 		disk->fops = &nbd_fops;
781 		disk->private_data = &nbd_dev[i];
782 		sprintf(disk->disk_name, "nbd%d", i);
783 		set_capacity(disk, 0);
784 		add_disk(disk);
785 	}
786 
787 	return 0;
788 out:
789 	while (i--) {
790 		blk_cleanup_queue(nbd_dev[i].disk->queue);
791 		put_disk(nbd_dev[i].disk);
792 	}
793 	kfree(nbd_dev);
794 	return err;
795 }
796 
797 static void __exit nbd_cleanup(void)
798 {
799 	int i;
800 	for (i = 0; i < nbds_max; i++) {
801 		struct gendisk *disk = nbd_dev[i].disk;
802 		nbd_dev[i].magic = 0;
803 		if (disk) {
804 			del_gendisk(disk);
805 			blk_cleanup_queue(disk->queue);
806 			put_disk(disk);
807 		}
808 	}
809 	unregister_blkdev(NBD_MAJOR, "nbd");
810 	kfree(nbd_dev);
811 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
812 }
813 
814 module_init(nbd_init);
815 module_exit(nbd_cleanup);
816 
817 MODULE_DESCRIPTION("Network Block Device");
818 MODULE_LICENSE("GPL");
819 
820 module_param(nbds_max, int, 0444);
821 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
822 module_param(max_part, int, 0444);
823 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
824 #ifndef NDEBUG
825 module_param(debugflags, int, 0644);
826 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
827 #endif
828