xref: /openbmc/linux/drivers/block/nbd.c (revision 95e9fd10)
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 
36 #include <asm/uaccess.h>
37 #include <asm/types.h>
38 
39 #include <linux/nbd.h>
40 
41 #define NBD_MAGIC 0x68797548
42 
43 #ifdef NDEBUG
44 #define dprintk(flags, fmt...)
45 #else /* NDEBUG */
46 #define dprintk(flags, fmt...) do { \
47 	if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
48 } while (0)
49 #define DBG_IOCTL       0x0004
50 #define DBG_INIT        0x0010
51 #define DBG_EXIT        0x0020
52 #define DBG_BLKDEV      0x0100
53 #define DBG_RX          0x0200
54 #define DBG_TX          0x0400
55 static unsigned int debugflags;
56 #endif /* NDEBUG */
57 
58 static unsigned int nbds_max = 16;
59 static struct nbd_device *nbd_dev;
60 static int max_part;
61 
62 /*
63  * Use just one lock (or at most 1 per NIC). Two arguments for this:
64  * 1. Each NIC is essentially a synchronization point for all servers
65  *    accessed through that NIC so there's no need to have more locks
66  *    than NICs anyway.
67  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
68  *    down each lock to the point where they're actually slower than just
69  *    a single lock.
70  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
71  */
72 static DEFINE_SPINLOCK(nbd_lock);
73 
74 #ifndef NDEBUG
75 static const char *ioctl_cmd_to_ascii(int cmd)
76 {
77 	switch (cmd) {
78 	case NBD_SET_SOCK: return "set-sock";
79 	case NBD_SET_BLKSIZE: return "set-blksize";
80 	case NBD_SET_SIZE: return "set-size";
81 	case NBD_DO_IT: return "do-it";
82 	case NBD_CLEAR_SOCK: return "clear-sock";
83 	case NBD_CLEAR_QUE: return "clear-que";
84 	case NBD_PRINT_DEBUG: return "print-debug";
85 	case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
86 	case NBD_DISCONNECT: return "disconnect";
87 	case BLKROSET: return "set-read-only";
88 	case BLKFLSBUF: return "flush-buffer-cache";
89 	}
90 	return "unknown";
91 }
92 
93 static const char *nbdcmd_to_ascii(int cmd)
94 {
95 	switch (cmd) {
96 	case  NBD_CMD_READ: return "read";
97 	case NBD_CMD_WRITE: return "write";
98 	case  NBD_CMD_DISC: return "disconnect";
99 	}
100 	return "invalid";
101 }
102 #endif /* NDEBUG */
103 
104 static void nbd_end_request(struct request *req)
105 {
106 	int error = req->errors ? -EIO : 0;
107 	struct request_queue *q = req->q;
108 	unsigned long flags;
109 
110 	dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
111 			req, error ? "failed" : "done");
112 
113 	spin_lock_irqsave(q->queue_lock, flags);
114 	__blk_end_request_all(req, error);
115 	spin_unlock_irqrestore(q->queue_lock, flags);
116 }
117 
118 static void sock_shutdown(struct nbd_device *nbd, int lock)
119 {
120 	/* Forcibly shutdown the socket causing all listeners
121 	 * to error
122 	 *
123 	 * FIXME: This code is duplicated from sys_shutdown, but
124 	 * there should be a more generic interface rather than
125 	 * calling socket ops directly here */
126 	if (lock)
127 		mutex_lock(&nbd->tx_lock);
128 	if (nbd->sock) {
129 		dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
130 		kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
131 		nbd->sock = NULL;
132 	}
133 	if (lock)
134 		mutex_unlock(&nbd->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 *nbd, int send, void *buf, int size,
150 		int msg_flags)
151 {
152 	struct socket *sock = nbd->sock;
153 	int result;
154 	struct msghdr msg;
155 	struct kvec iov;
156 	sigset_t blocked, oldset;
157 	unsigned long pflags = current->flags;
158 
159 	if (unlikely(!sock)) {
160 		dev_err(disk_to_dev(nbd->disk),
161 			"Attempted %s on closed socket in sock_xmit\n",
162 			(send ? "send" : "recv"));
163 		return -EINVAL;
164 	}
165 
166 	/* Allow interception of SIGKILL only
167 	 * Don't allow other signals to interrupt the transmission */
168 	siginitsetinv(&blocked, sigmask(SIGKILL));
169 	sigprocmask(SIG_SETMASK, &blocked, &oldset);
170 
171 	current->flags |= PF_MEMALLOC;
172 	do {
173 		sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
174 		iov.iov_base = buf;
175 		iov.iov_len = size;
176 		msg.msg_name = NULL;
177 		msg.msg_namelen = 0;
178 		msg.msg_control = NULL;
179 		msg.msg_controllen = 0;
180 		msg.msg_flags = msg_flags | MSG_NOSIGNAL;
181 
182 		if (send) {
183 			struct timer_list ti;
184 
185 			if (nbd->xmit_timeout) {
186 				init_timer(&ti);
187 				ti.function = nbd_xmit_timeout;
188 				ti.data = (unsigned long)current;
189 				ti.expires = jiffies + nbd->xmit_timeout;
190 				add_timer(&ti);
191 			}
192 			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
193 			if (nbd->xmit_timeout)
194 				del_timer_sync(&ti);
195 		} else
196 			result = kernel_recvmsg(sock, &msg, &iov, 1, size,
197 						msg.msg_flags);
198 
199 		if (signal_pending(current)) {
200 			siginfo_t info;
201 			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
202 				task_pid_nr(current), current->comm,
203 				dequeue_signal_lock(current, &current->blocked, &info));
204 			result = -EINTR;
205 			sock_shutdown(nbd, !send);
206 			break;
207 		}
208 
209 		if (result <= 0) {
210 			if (result == 0)
211 				result = -EPIPE; /* short read */
212 			break;
213 		}
214 		size -= result;
215 		buf += result;
216 	} while (size > 0);
217 
218 	sigprocmask(SIG_SETMASK, &oldset, NULL);
219 	tsk_restore_flags(current, pflags, PF_MEMALLOC);
220 
221 	return result;
222 }
223 
224 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
225 		int flags)
226 {
227 	int result;
228 	void *kaddr = kmap(bvec->bv_page);
229 	result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
230 			   bvec->bv_len, flags);
231 	kunmap(bvec->bv_page);
232 	return result;
233 }
234 
235 /* always call with the tx_lock held */
236 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
237 {
238 	int result, flags;
239 	struct nbd_request request;
240 	unsigned long size = blk_rq_bytes(req);
241 
242 	request.magic = htonl(NBD_REQUEST_MAGIC);
243 	request.type = htonl(nbd_cmd(req));
244 	request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
245 	request.len = htonl(size);
246 	memcpy(request.handle, &req, sizeof(req));
247 
248 	dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
249 			nbd->disk->disk_name, req,
250 			nbdcmd_to_ascii(nbd_cmd(req)),
251 			(unsigned long long)blk_rq_pos(req) << 9,
252 			blk_rq_bytes(req));
253 	result = sock_xmit(nbd, 1, &request, sizeof(request),
254 			(nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
255 	if (result <= 0) {
256 		dev_err(disk_to_dev(nbd->disk),
257 			"Send control failed (result %d)\n", result);
258 		goto error_out;
259 	}
260 
261 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
262 		struct req_iterator iter;
263 		struct bio_vec *bvec;
264 		/*
265 		 * we are really probing at internals to determine
266 		 * whether to set MSG_MORE or not...
267 		 */
268 		rq_for_each_segment(bvec, req, iter) {
269 			flags = 0;
270 			if (!rq_iter_last(req, iter))
271 				flags = MSG_MORE;
272 			dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
273 					nbd->disk->disk_name, req, bvec->bv_len);
274 			result = sock_send_bvec(nbd, bvec, flags);
275 			if (result <= 0) {
276 				dev_err(disk_to_dev(nbd->disk),
277 					"Send data failed (result %d)\n",
278 					result);
279 				goto error_out;
280 			}
281 		}
282 	}
283 	return 0;
284 
285 error_out:
286 	return -EIO;
287 }
288 
289 static struct request *nbd_find_request(struct nbd_device *nbd,
290 					struct request *xreq)
291 {
292 	struct request *req, *tmp;
293 	int err;
294 
295 	err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
296 	if (unlikely(err))
297 		goto out;
298 
299 	spin_lock(&nbd->queue_lock);
300 	list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
301 		if (req != xreq)
302 			continue;
303 		list_del_init(&req->queuelist);
304 		spin_unlock(&nbd->queue_lock);
305 		return req;
306 	}
307 	spin_unlock(&nbd->queue_lock);
308 
309 	err = -ENOENT;
310 
311 out:
312 	return ERR_PTR(err);
313 }
314 
315 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
316 {
317 	int result;
318 	void *kaddr = kmap(bvec->bv_page);
319 	result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
320 			MSG_WAITALL);
321 	kunmap(bvec->bv_page);
322 	return result;
323 }
324 
325 /* NULL returned = something went wrong, inform userspace */
326 static struct request *nbd_read_stat(struct nbd_device *nbd)
327 {
328 	int result;
329 	struct nbd_reply reply;
330 	struct request *req;
331 
332 	reply.magic = 0;
333 	result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
334 	if (result <= 0) {
335 		dev_err(disk_to_dev(nbd->disk),
336 			"Receive control failed (result %d)\n", result);
337 		goto harderror;
338 	}
339 
340 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
341 		dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
342 				(unsigned long)ntohl(reply.magic));
343 		result = -EPROTO;
344 		goto harderror;
345 	}
346 
347 	req = nbd_find_request(nbd, *(struct request **)reply.handle);
348 	if (IS_ERR(req)) {
349 		result = PTR_ERR(req);
350 		if (result != -ENOENT)
351 			goto harderror;
352 
353 		dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
354 			reply.handle);
355 		result = -EBADR;
356 		goto harderror;
357 	}
358 
359 	if (ntohl(reply.error)) {
360 		dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
361 			ntohl(reply.error));
362 		req->errors++;
363 		return req;
364 	}
365 
366 	dprintk(DBG_RX, "%s: request %p: got reply\n",
367 			nbd->disk->disk_name, req);
368 	if (nbd_cmd(req) == NBD_CMD_READ) {
369 		struct req_iterator iter;
370 		struct bio_vec *bvec;
371 
372 		rq_for_each_segment(bvec, req, iter) {
373 			result = sock_recv_bvec(nbd, bvec);
374 			if (result <= 0) {
375 				dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
376 					result);
377 				req->errors++;
378 				return req;
379 			}
380 			dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
381 				nbd->disk->disk_name, req, bvec->bv_len);
382 		}
383 	}
384 	return req;
385 harderror:
386 	nbd->harderror = result;
387 	return NULL;
388 }
389 
390 static ssize_t pid_show(struct device *dev,
391 			struct device_attribute *attr, char *buf)
392 {
393 	struct gendisk *disk = dev_to_disk(dev);
394 
395 	return sprintf(buf, "%ld\n",
396 		(long) ((struct nbd_device *)disk->private_data)->pid);
397 }
398 
399 static struct device_attribute pid_attr = {
400 	.attr = { .name = "pid", .mode = S_IRUGO},
401 	.show = pid_show,
402 };
403 
404 static int nbd_do_it(struct nbd_device *nbd)
405 {
406 	struct request *req;
407 	int ret;
408 
409 	BUG_ON(nbd->magic != NBD_MAGIC);
410 
411 	sk_set_memalloc(nbd->sock->sk);
412 	nbd->pid = task_pid_nr(current);
413 	ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
414 	if (ret) {
415 		dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
416 		nbd->pid = 0;
417 		return ret;
418 	}
419 
420 	while ((req = nbd_read_stat(nbd)) != NULL)
421 		nbd_end_request(req);
422 
423 	device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
424 	nbd->pid = 0;
425 	return 0;
426 }
427 
428 static void nbd_clear_que(struct nbd_device *nbd)
429 {
430 	struct request *req;
431 
432 	BUG_ON(nbd->magic != NBD_MAGIC);
433 
434 	/*
435 	 * Because we have set nbd->sock to NULL under the tx_lock, all
436 	 * modifications to the list must have completed by now.  For
437 	 * the same reason, the active_req must be NULL.
438 	 *
439 	 * As a consequence, we don't need to take the spin lock while
440 	 * purging the list here.
441 	 */
442 	BUG_ON(nbd->sock);
443 	BUG_ON(nbd->active_req);
444 
445 	while (!list_empty(&nbd->queue_head)) {
446 		req = list_entry(nbd->queue_head.next, struct request,
447 				 queuelist);
448 		list_del_init(&req->queuelist);
449 		req->errors++;
450 		nbd_end_request(req);
451 	}
452 }
453 
454 
455 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
456 {
457 	if (req->cmd_type != REQ_TYPE_FS)
458 		goto error_out;
459 
460 	nbd_cmd(req) = NBD_CMD_READ;
461 	if (rq_data_dir(req) == WRITE) {
462 		nbd_cmd(req) = NBD_CMD_WRITE;
463 		if (nbd->flags & NBD_READ_ONLY) {
464 			dev_err(disk_to_dev(nbd->disk),
465 				"Write on read-only\n");
466 			goto error_out;
467 		}
468 	}
469 
470 	req->errors = 0;
471 
472 	mutex_lock(&nbd->tx_lock);
473 	if (unlikely(!nbd->sock)) {
474 		mutex_unlock(&nbd->tx_lock);
475 		dev_err(disk_to_dev(nbd->disk),
476 			"Attempted send on closed socket\n");
477 		goto error_out;
478 	}
479 
480 	nbd->active_req = req;
481 
482 	if (nbd_send_req(nbd, req) != 0) {
483 		dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
484 		req->errors++;
485 		nbd_end_request(req);
486 	} else {
487 		spin_lock(&nbd->queue_lock);
488 		list_add_tail(&req->queuelist, &nbd->queue_head);
489 		spin_unlock(&nbd->queue_lock);
490 	}
491 
492 	nbd->active_req = NULL;
493 	mutex_unlock(&nbd->tx_lock);
494 	wake_up_all(&nbd->active_wq);
495 
496 	return;
497 
498 error_out:
499 	req->errors++;
500 	nbd_end_request(req);
501 }
502 
503 static int nbd_thread(void *data)
504 {
505 	struct nbd_device *nbd = data;
506 	struct request *req;
507 
508 	set_user_nice(current, -20);
509 	while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
510 		/* wait for something to do */
511 		wait_event_interruptible(nbd->waiting_wq,
512 					 kthread_should_stop() ||
513 					 !list_empty(&nbd->waiting_queue));
514 
515 		/* extract request */
516 		if (list_empty(&nbd->waiting_queue))
517 			continue;
518 
519 		spin_lock_irq(&nbd->queue_lock);
520 		req = list_entry(nbd->waiting_queue.next, struct request,
521 				 queuelist);
522 		list_del_init(&req->queuelist);
523 		spin_unlock_irq(&nbd->queue_lock);
524 
525 		/* handle request */
526 		nbd_handle_req(nbd, req);
527 	}
528 	return 0;
529 }
530 
531 /*
532  * We always wait for result of write, for now. It would be nice to make it optional
533  * in future
534  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
535  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
536  */
537 
538 static void do_nbd_request(struct request_queue *q)
539 {
540 	struct request *req;
541 
542 	while ((req = blk_fetch_request(q)) != NULL) {
543 		struct nbd_device *nbd;
544 
545 		spin_unlock_irq(q->queue_lock);
546 
547 		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
548 				req->rq_disk->disk_name, req, req->cmd_type);
549 
550 		nbd = req->rq_disk->private_data;
551 
552 		BUG_ON(nbd->magic != NBD_MAGIC);
553 
554 		if (unlikely(!nbd->sock)) {
555 			dev_err(disk_to_dev(nbd->disk),
556 				"Attempted send on closed socket\n");
557 			req->errors++;
558 			nbd_end_request(req);
559 			spin_lock_irq(q->queue_lock);
560 			continue;
561 		}
562 
563 		spin_lock_irq(&nbd->queue_lock);
564 		list_add_tail(&req->queuelist, &nbd->waiting_queue);
565 		spin_unlock_irq(&nbd->queue_lock);
566 
567 		wake_up(&nbd->waiting_wq);
568 
569 		spin_lock_irq(q->queue_lock);
570 	}
571 }
572 
573 /* Must be called with tx_lock held */
574 
575 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
576 		       unsigned int cmd, unsigned long arg)
577 {
578 	switch (cmd) {
579 	case NBD_DISCONNECT: {
580 		struct request sreq;
581 
582 		dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
583 
584 		blk_rq_init(NULL, &sreq);
585 		sreq.cmd_type = REQ_TYPE_SPECIAL;
586 		nbd_cmd(&sreq) = NBD_CMD_DISC;
587 		if (!nbd->sock)
588 			return -EINVAL;
589 		nbd_send_req(nbd, &sreq);
590                 return 0;
591 	}
592 
593 	case NBD_CLEAR_SOCK: {
594 		struct file *file;
595 
596 		nbd->sock = NULL;
597 		file = nbd->file;
598 		nbd->file = NULL;
599 		nbd_clear_que(nbd);
600 		BUG_ON(!list_empty(&nbd->queue_head));
601 		if (file)
602 			fput(file);
603 		return 0;
604 	}
605 
606 	case NBD_SET_SOCK: {
607 		struct file *file;
608 		if (nbd->file)
609 			return -EBUSY;
610 		file = fget(arg);
611 		if (file) {
612 			struct inode *inode = file->f_path.dentry->d_inode;
613 			if (S_ISSOCK(inode->i_mode)) {
614 				nbd->file = file;
615 				nbd->sock = SOCKET_I(inode);
616 				if (max_part > 0)
617 					bdev->bd_invalidated = 1;
618 				return 0;
619 			} else {
620 				fput(file);
621 			}
622 		}
623 		return -EINVAL;
624 	}
625 
626 	case NBD_SET_BLKSIZE:
627 		nbd->blksize = arg;
628 		nbd->bytesize &= ~(nbd->blksize-1);
629 		bdev->bd_inode->i_size = nbd->bytesize;
630 		set_blocksize(bdev, nbd->blksize);
631 		set_capacity(nbd->disk, nbd->bytesize >> 9);
632 		return 0;
633 
634 	case NBD_SET_SIZE:
635 		nbd->bytesize = arg & ~(nbd->blksize-1);
636 		bdev->bd_inode->i_size = nbd->bytesize;
637 		set_blocksize(bdev, nbd->blksize);
638 		set_capacity(nbd->disk, nbd->bytesize >> 9);
639 		return 0;
640 
641 	case NBD_SET_TIMEOUT:
642 		nbd->xmit_timeout = arg * HZ;
643 		return 0;
644 
645 	case NBD_SET_SIZE_BLOCKS:
646 		nbd->bytesize = ((u64) arg) * nbd->blksize;
647 		bdev->bd_inode->i_size = nbd->bytesize;
648 		set_blocksize(bdev, nbd->blksize);
649 		set_capacity(nbd->disk, nbd->bytesize >> 9);
650 		return 0;
651 
652 	case NBD_DO_IT: {
653 		struct task_struct *thread;
654 		struct file *file;
655 		int error;
656 
657 		if (nbd->pid)
658 			return -EBUSY;
659 		if (!nbd->file)
660 			return -EINVAL;
661 
662 		mutex_unlock(&nbd->tx_lock);
663 
664 		thread = kthread_create(nbd_thread, nbd, nbd->disk->disk_name);
665 		if (IS_ERR(thread)) {
666 			mutex_lock(&nbd->tx_lock);
667 			return PTR_ERR(thread);
668 		}
669 		wake_up_process(thread);
670 		error = nbd_do_it(nbd);
671 		kthread_stop(thread);
672 
673 		mutex_lock(&nbd->tx_lock);
674 		if (error)
675 			return error;
676 		sock_shutdown(nbd, 0);
677 		file = nbd->file;
678 		nbd->file = NULL;
679 		nbd_clear_que(nbd);
680 		dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
681 		if (file)
682 			fput(file);
683 		nbd->bytesize = 0;
684 		bdev->bd_inode->i_size = 0;
685 		set_capacity(nbd->disk, 0);
686 		if (max_part > 0)
687 			ioctl_by_bdev(bdev, BLKRRPART, 0);
688 		return nbd->harderror;
689 	}
690 
691 	case NBD_CLEAR_QUE:
692 		/*
693 		 * This is for compatibility only.  The queue is always cleared
694 		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
695 		 */
696 		BUG_ON(!nbd->sock && !list_empty(&nbd->queue_head));
697 		return 0;
698 
699 	case NBD_PRINT_DEBUG:
700 		dev_info(disk_to_dev(nbd->disk),
701 			"next = %p, prev = %p, head = %p\n",
702 			nbd->queue_head.next, nbd->queue_head.prev,
703 			&nbd->queue_head);
704 		return 0;
705 	}
706 	return -ENOTTY;
707 }
708 
709 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
710 		     unsigned int cmd, unsigned long arg)
711 {
712 	struct nbd_device *nbd = bdev->bd_disk->private_data;
713 	int error;
714 
715 	if (!capable(CAP_SYS_ADMIN))
716 		return -EPERM;
717 
718 	BUG_ON(nbd->magic != NBD_MAGIC);
719 
720 	/* Anyone capable of this syscall can do *real bad* things */
721 	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
722 		nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
723 
724 	mutex_lock(&nbd->tx_lock);
725 	error = __nbd_ioctl(bdev, nbd, cmd, arg);
726 	mutex_unlock(&nbd->tx_lock);
727 
728 	return error;
729 }
730 
731 static const struct block_device_operations nbd_fops =
732 {
733 	.owner =	THIS_MODULE,
734 	.ioctl =	nbd_ioctl,
735 };
736 
737 /*
738  * And here should be modules and kernel interface
739  *  (Just smiley confuses emacs :-)
740  */
741 
742 static int __init nbd_init(void)
743 {
744 	int err = -ENOMEM;
745 	int i;
746 	int part_shift;
747 
748 	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
749 
750 	if (max_part < 0) {
751 		printk(KERN_ERR "nbd: max_part must be >= 0\n");
752 		return -EINVAL;
753 	}
754 
755 	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
756 	if (!nbd_dev)
757 		return -ENOMEM;
758 
759 	part_shift = 0;
760 	if (max_part > 0) {
761 		part_shift = fls(max_part);
762 
763 		/*
764 		 * Adjust max_part according to part_shift as it is exported
765 		 * to user space so that user can know the max number of
766 		 * partition kernel should be able to manage.
767 		 *
768 		 * Note that -1 is required because partition 0 is reserved
769 		 * for the whole disk.
770 		 */
771 		max_part = (1UL << part_shift) - 1;
772 	}
773 
774 	if ((1UL << part_shift) > DISK_MAX_PARTS)
775 		return -EINVAL;
776 
777 	if (nbds_max > 1UL << (MINORBITS - part_shift))
778 		return -EINVAL;
779 
780 	for (i = 0; i < nbds_max; i++) {
781 		struct gendisk *disk = alloc_disk(1 << part_shift);
782 		if (!disk)
783 			goto out;
784 		nbd_dev[i].disk = disk;
785 		/*
786 		 * The new linux 2.5 block layer implementation requires
787 		 * every gendisk to have its very own request_queue struct.
788 		 * These structs are big so we dynamically allocate them.
789 		 */
790 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
791 		if (!disk->queue) {
792 			put_disk(disk);
793 			goto out;
794 		}
795 		/*
796 		 * Tell the block layer that we are not a rotational device
797 		 */
798 		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
799 	}
800 
801 	if (register_blkdev(NBD_MAJOR, "nbd")) {
802 		err = -EIO;
803 		goto out;
804 	}
805 
806 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
807 	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
808 
809 	for (i = 0; i < nbds_max; i++) {
810 		struct gendisk *disk = nbd_dev[i].disk;
811 		nbd_dev[i].file = NULL;
812 		nbd_dev[i].magic = NBD_MAGIC;
813 		nbd_dev[i].flags = 0;
814 		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
815 		spin_lock_init(&nbd_dev[i].queue_lock);
816 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
817 		mutex_init(&nbd_dev[i].tx_lock);
818 		init_waitqueue_head(&nbd_dev[i].active_wq);
819 		init_waitqueue_head(&nbd_dev[i].waiting_wq);
820 		nbd_dev[i].blksize = 1024;
821 		nbd_dev[i].bytesize = 0;
822 		disk->major = NBD_MAJOR;
823 		disk->first_minor = i << part_shift;
824 		disk->fops = &nbd_fops;
825 		disk->private_data = &nbd_dev[i];
826 		sprintf(disk->disk_name, "nbd%d", i);
827 		set_capacity(disk, 0);
828 		add_disk(disk);
829 	}
830 
831 	return 0;
832 out:
833 	while (i--) {
834 		blk_cleanup_queue(nbd_dev[i].disk->queue);
835 		put_disk(nbd_dev[i].disk);
836 	}
837 	kfree(nbd_dev);
838 	return err;
839 }
840 
841 static void __exit nbd_cleanup(void)
842 {
843 	int i;
844 	for (i = 0; i < nbds_max; i++) {
845 		struct gendisk *disk = nbd_dev[i].disk;
846 		nbd_dev[i].magic = 0;
847 		if (disk) {
848 			del_gendisk(disk);
849 			blk_cleanup_queue(disk->queue);
850 			put_disk(disk);
851 		}
852 	}
853 	unregister_blkdev(NBD_MAJOR, "nbd");
854 	kfree(nbd_dev);
855 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
856 }
857 
858 module_init(nbd_init);
859 module_exit(nbd_cleanup);
860 
861 MODULE_DESCRIPTION("Network Block Device");
862 MODULE_LICENSE("GPL");
863 
864 module_param(nbds_max, int, 0444);
865 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
866 module_param(max_part, int, 0444);
867 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
868 #ifndef NDEBUG
869 module_param(debugflags, int, 0644);
870 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
871 #endif
872