xref: /openbmc/linux/drivers/block/nbd.c (revision 9f380456)
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 
158 	if (unlikely(!sock)) {
159 		dev_err(disk_to_dev(nbd->disk),
160 			"Attempted %s on closed socket in sock_xmit\n",
161 			(send ? "send" : "recv"));
162 		return -EINVAL;
163 	}
164 
165 	/* Allow interception of SIGKILL only
166 	 * Don't allow other signals to interrupt the transmission */
167 	siginitsetinv(&blocked, sigmask(SIGKILL));
168 	sigprocmask(SIG_SETMASK, &blocked, &oldset);
169 
170 	do {
171 		sock->sk->sk_allocation = GFP_NOIO;
172 		iov.iov_base = buf;
173 		iov.iov_len = size;
174 		msg.msg_name = NULL;
175 		msg.msg_namelen = 0;
176 		msg.msg_control = NULL;
177 		msg.msg_controllen = 0;
178 		msg.msg_flags = msg_flags | MSG_NOSIGNAL;
179 
180 		if (send) {
181 			struct timer_list ti;
182 
183 			if (nbd->xmit_timeout) {
184 				init_timer(&ti);
185 				ti.function = nbd_xmit_timeout;
186 				ti.data = (unsigned long)current;
187 				ti.expires = jiffies + nbd->xmit_timeout;
188 				add_timer(&ti);
189 			}
190 			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
191 			if (nbd->xmit_timeout)
192 				del_timer_sync(&ti);
193 		} else
194 			result = kernel_recvmsg(sock, &msg, &iov, 1, size,
195 						msg.msg_flags);
196 
197 		if (signal_pending(current)) {
198 			siginfo_t info;
199 			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
200 				task_pid_nr(current), current->comm,
201 				dequeue_signal_lock(current, &current->blocked, &info));
202 			result = -EINTR;
203 			sock_shutdown(nbd, !send);
204 			break;
205 		}
206 
207 		if (result <= 0) {
208 			if (result == 0)
209 				result = -EPIPE; /* short read */
210 			break;
211 		}
212 		size -= result;
213 		buf += result;
214 	} while (size > 0);
215 
216 	sigprocmask(SIG_SETMASK, &oldset, NULL);
217 
218 	return result;
219 }
220 
221 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
222 		int flags)
223 {
224 	int result;
225 	void *kaddr = kmap(bvec->bv_page);
226 	result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
227 			   bvec->bv_len, flags);
228 	kunmap(bvec->bv_page);
229 	return result;
230 }
231 
232 /* always call with the tx_lock held */
233 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
234 {
235 	int result, flags;
236 	struct nbd_request request;
237 	unsigned long size = blk_rq_bytes(req);
238 
239 	request.magic = htonl(NBD_REQUEST_MAGIC);
240 	request.type = htonl(nbd_cmd(req));
241 	request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
242 	request.len = htonl(size);
243 	memcpy(request.handle, &req, sizeof(req));
244 
245 	dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
246 			nbd->disk->disk_name, req,
247 			nbdcmd_to_ascii(nbd_cmd(req)),
248 			(unsigned long long)blk_rq_pos(req) << 9,
249 			blk_rq_bytes(req));
250 	result = sock_xmit(nbd, 1, &request, sizeof(request),
251 			(nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
252 	if (result <= 0) {
253 		dev_err(disk_to_dev(nbd->disk),
254 			"Send control failed (result %d)\n", result);
255 		goto error_out;
256 	}
257 
258 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
259 		struct req_iterator iter;
260 		struct bio_vec *bvec;
261 		/*
262 		 * we are really probing at internals to determine
263 		 * whether to set MSG_MORE or not...
264 		 */
265 		rq_for_each_segment(bvec, req, iter) {
266 			flags = 0;
267 			if (!rq_iter_last(req, iter))
268 				flags = MSG_MORE;
269 			dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
270 					nbd->disk->disk_name, req, bvec->bv_len);
271 			result = sock_send_bvec(nbd, bvec, flags);
272 			if (result <= 0) {
273 				dev_err(disk_to_dev(nbd->disk),
274 					"Send data failed (result %d)\n",
275 					result);
276 				goto error_out;
277 			}
278 		}
279 	}
280 	return 0;
281 
282 error_out:
283 	return -EIO;
284 }
285 
286 static struct request *nbd_find_request(struct nbd_device *nbd,
287 					struct request *xreq)
288 {
289 	struct request *req, *tmp;
290 	int err;
291 
292 	err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
293 	if (unlikely(err))
294 		goto out;
295 
296 	spin_lock(&nbd->queue_lock);
297 	list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
298 		if (req != xreq)
299 			continue;
300 		list_del_init(&req->queuelist);
301 		spin_unlock(&nbd->queue_lock);
302 		return req;
303 	}
304 	spin_unlock(&nbd->queue_lock);
305 
306 	err = -ENOENT;
307 
308 out:
309 	return ERR_PTR(err);
310 }
311 
312 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
313 {
314 	int result;
315 	void *kaddr = kmap(bvec->bv_page);
316 	result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
317 			MSG_WAITALL);
318 	kunmap(bvec->bv_page);
319 	return result;
320 }
321 
322 /* NULL returned = something went wrong, inform userspace */
323 static struct request *nbd_read_stat(struct nbd_device *nbd)
324 {
325 	int result;
326 	struct nbd_reply reply;
327 	struct request *req;
328 
329 	reply.magic = 0;
330 	result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
331 	if (result <= 0) {
332 		dev_err(disk_to_dev(nbd->disk),
333 			"Receive control failed (result %d)\n", result);
334 		goto harderror;
335 	}
336 
337 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
338 		dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
339 				(unsigned long)ntohl(reply.magic));
340 		result = -EPROTO;
341 		goto harderror;
342 	}
343 
344 	req = nbd_find_request(nbd, *(struct request **)reply.handle);
345 	if (IS_ERR(req)) {
346 		result = PTR_ERR(req);
347 		if (result != -ENOENT)
348 			goto harderror;
349 
350 		dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
351 			reply.handle);
352 		result = -EBADR;
353 		goto harderror;
354 	}
355 
356 	if (ntohl(reply.error)) {
357 		dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
358 			ntohl(reply.error));
359 		req->errors++;
360 		return req;
361 	}
362 
363 	dprintk(DBG_RX, "%s: request %p: got reply\n",
364 			nbd->disk->disk_name, req);
365 	if (nbd_cmd(req) == NBD_CMD_READ) {
366 		struct req_iterator iter;
367 		struct bio_vec *bvec;
368 
369 		rq_for_each_segment(bvec, req, iter) {
370 			result = sock_recv_bvec(nbd, bvec);
371 			if (result <= 0) {
372 				dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
373 					result);
374 				req->errors++;
375 				return req;
376 			}
377 			dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
378 				nbd->disk->disk_name, req, bvec->bv_len);
379 		}
380 	}
381 	return req;
382 harderror:
383 	nbd->harderror = result;
384 	return NULL;
385 }
386 
387 static ssize_t pid_show(struct device *dev,
388 			struct device_attribute *attr, char *buf)
389 {
390 	struct gendisk *disk = dev_to_disk(dev);
391 
392 	return sprintf(buf, "%ld\n",
393 		(long) ((struct nbd_device *)disk->private_data)->pid);
394 }
395 
396 static struct device_attribute pid_attr = {
397 	.attr = { .name = "pid", .mode = S_IRUGO},
398 	.show = pid_show,
399 };
400 
401 static int nbd_do_it(struct nbd_device *nbd)
402 {
403 	struct request *req;
404 	int ret;
405 
406 	BUG_ON(nbd->magic != NBD_MAGIC);
407 
408 	nbd->pid = task_pid_nr(current);
409 	ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
410 	if (ret) {
411 		dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
412 		nbd->pid = 0;
413 		return ret;
414 	}
415 
416 	while ((req = nbd_read_stat(nbd)) != NULL)
417 		nbd_end_request(req);
418 
419 	device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
420 	nbd->pid = 0;
421 	return 0;
422 }
423 
424 static void nbd_clear_que(struct nbd_device *nbd)
425 {
426 	struct request *req;
427 
428 	BUG_ON(nbd->magic != NBD_MAGIC);
429 
430 	/*
431 	 * Because we have set nbd->sock to NULL under the tx_lock, all
432 	 * modifications to the list must have completed by now.  For
433 	 * the same reason, the active_req must be NULL.
434 	 *
435 	 * As a consequence, we don't need to take the spin lock while
436 	 * purging the list here.
437 	 */
438 	BUG_ON(nbd->sock);
439 	BUG_ON(nbd->active_req);
440 
441 	while (!list_empty(&nbd->queue_head)) {
442 		req = list_entry(nbd->queue_head.next, struct request,
443 				 queuelist);
444 		list_del_init(&req->queuelist);
445 		req->errors++;
446 		nbd_end_request(req);
447 	}
448 }
449 
450 
451 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
452 {
453 	if (req->cmd_type != REQ_TYPE_FS)
454 		goto error_out;
455 
456 	nbd_cmd(req) = NBD_CMD_READ;
457 	if (rq_data_dir(req) == WRITE) {
458 		nbd_cmd(req) = NBD_CMD_WRITE;
459 		if (nbd->flags & NBD_READ_ONLY) {
460 			dev_err(disk_to_dev(nbd->disk),
461 				"Write on read-only\n");
462 			goto error_out;
463 		}
464 	}
465 
466 	req->errors = 0;
467 
468 	mutex_lock(&nbd->tx_lock);
469 	if (unlikely(!nbd->sock)) {
470 		mutex_unlock(&nbd->tx_lock);
471 		dev_err(disk_to_dev(nbd->disk),
472 			"Attempted send on closed socket\n");
473 		goto error_out;
474 	}
475 
476 	nbd->active_req = req;
477 
478 	if (nbd_send_req(nbd, req) != 0) {
479 		dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
480 		req->errors++;
481 		nbd_end_request(req);
482 	} else {
483 		spin_lock(&nbd->queue_lock);
484 		list_add(&req->queuelist, &nbd->queue_head);
485 		spin_unlock(&nbd->queue_lock);
486 	}
487 
488 	nbd->active_req = NULL;
489 	mutex_unlock(&nbd->tx_lock);
490 	wake_up_all(&nbd->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 *nbd = data;
502 	struct request *req;
503 
504 	set_user_nice(current, -20);
505 	while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
506 		/* wait for something to do */
507 		wait_event_interruptible(nbd->waiting_wq,
508 					 kthread_should_stop() ||
509 					 !list_empty(&nbd->waiting_queue));
510 
511 		/* extract request */
512 		if (list_empty(&nbd->waiting_queue))
513 			continue;
514 
515 		spin_lock_irq(&nbd->queue_lock);
516 		req = list_entry(nbd->waiting_queue.next, struct request,
517 				 queuelist);
518 		list_del_init(&req->queuelist);
519 		spin_unlock_irq(&nbd->queue_lock);
520 
521 		/* handle request */
522 		nbd_handle_req(nbd, 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) && (nbd->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 = blk_fetch_request(q)) != NULL) {
539 		struct nbd_device *nbd;
540 
541 		spin_unlock_irq(q->queue_lock);
542 
543 		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
544 				req->rq_disk->disk_name, req, req->cmd_type);
545 
546 		nbd = req->rq_disk->private_data;
547 
548 		BUG_ON(nbd->magic != NBD_MAGIC);
549 
550 		if (unlikely(!nbd->sock)) {
551 			dev_err(disk_to_dev(nbd->disk),
552 				"Attempted send on closed socket\n");
553 			req->errors++;
554 			nbd_end_request(req);
555 			spin_lock_irq(q->queue_lock);
556 			continue;
557 		}
558 
559 		spin_lock_irq(&nbd->queue_lock);
560 		list_add_tail(&req->queuelist, &nbd->waiting_queue);
561 		spin_unlock_irq(&nbd->queue_lock);
562 
563 		wake_up(&nbd->waiting_wq);
564 
565 		spin_lock_irq(q->queue_lock);
566 	}
567 }
568 
569 /* Must be called with tx_lock held */
570 
571 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
572 		       unsigned int cmd, unsigned long arg)
573 {
574 	switch (cmd) {
575 	case NBD_DISCONNECT: {
576 		struct request sreq;
577 
578 		dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
579 
580 		blk_rq_init(NULL, &sreq);
581 		sreq.cmd_type = REQ_TYPE_SPECIAL;
582 		nbd_cmd(&sreq) = NBD_CMD_DISC;
583 		if (!nbd->sock)
584 			return -EINVAL;
585 		nbd_send_req(nbd, &sreq);
586                 return 0;
587 	}
588 
589 	case NBD_CLEAR_SOCK: {
590 		struct file *file;
591 
592 		nbd->sock = NULL;
593 		file = nbd->file;
594 		nbd->file = NULL;
595 		nbd_clear_que(nbd);
596 		BUG_ON(!list_empty(&nbd->queue_head));
597 		if (file)
598 			fput(file);
599 		return 0;
600 	}
601 
602 	case NBD_SET_SOCK: {
603 		struct file *file;
604 		if (nbd->file)
605 			return -EBUSY;
606 		file = fget(arg);
607 		if (file) {
608 			struct inode *inode = file->f_path.dentry->d_inode;
609 			if (S_ISSOCK(inode->i_mode)) {
610 				nbd->file = file;
611 				nbd->sock = SOCKET_I(inode);
612 				if (max_part > 0)
613 					bdev->bd_invalidated = 1;
614 				return 0;
615 			} else {
616 				fput(file);
617 			}
618 		}
619 		return -EINVAL;
620 	}
621 
622 	case NBD_SET_BLKSIZE:
623 		nbd->blksize = arg;
624 		nbd->bytesize &= ~(nbd->blksize-1);
625 		bdev->bd_inode->i_size = nbd->bytesize;
626 		set_blocksize(bdev, nbd->blksize);
627 		set_capacity(nbd->disk, nbd->bytesize >> 9);
628 		return 0;
629 
630 	case NBD_SET_SIZE:
631 		nbd->bytesize = arg & ~(nbd->blksize-1);
632 		bdev->bd_inode->i_size = nbd->bytesize;
633 		set_blocksize(bdev, nbd->blksize);
634 		set_capacity(nbd->disk, nbd->bytesize >> 9);
635 		return 0;
636 
637 	case NBD_SET_TIMEOUT:
638 		nbd->xmit_timeout = arg * HZ;
639 		return 0;
640 
641 	case NBD_SET_SIZE_BLOCKS:
642 		nbd->bytesize = ((u64) arg) * nbd->blksize;
643 		bdev->bd_inode->i_size = nbd->bytesize;
644 		set_blocksize(bdev, nbd->blksize);
645 		set_capacity(nbd->disk, nbd->bytesize >> 9);
646 		return 0;
647 
648 	case NBD_DO_IT: {
649 		struct task_struct *thread;
650 		struct file *file;
651 		int error;
652 
653 		if (nbd->pid)
654 			return -EBUSY;
655 		if (!nbd->file)
656 			return -EINVAL;
657 
658 		mutex_unlock(&nbd->tx_lock);
659 
660 		thread = kthread_create(nbd_thread, nbd, nbd->disk->disk_name);
661 		if (IS_ERR(thread)) {
662 			mutex_lock(&nbd->tx_lock);
663 			return PTR_ERR(thread);
664 		}
665 		wake_up_process(thread);
666 		error = nbd_do_it(nbd);
667 		kthread_stop(thread);
668 
669 		mutex_lock(&nbd->tx_lock);
670 		if (error)
671 			return error;
672 		sock_shutdown(nbd, 0);
673 		file = nbd->file;
674 		nbd->file = NULL;
675 		nbd_clear_que(nbd);
676 		dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
677 		if (file)
678 			fput(file);
679 		nbd->bytesize = 0;
680 		bdev->bd_inode->i_size = 0;
681 		set_capacity(nbd->disk, 0);
682 		if (max_part > 0)
683 			ioctl_by_bdev(bdev, BLKRRPART, 0);
684 		return nbd->harderror;
685 	}
686 
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(!nbd->sock && !list_empty(&nbd->queue_head));
693 		return 0;
694 
695 	case NBD_PRINT_DEBUG:
696 		dev_info(disk_to_dev(nbd->disk),
697 			"next = %p, prev = %p, head = %p\n",
698 			nbd->queue_head.next, nbd->queue_head.prev,
699 			&nbd->queue_head);
700 		return 0;
701 	}
702 	return -ENOTTY;
703 }
704 
705 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
706 		     unsigned int cmd, unsigned long arg)
707 {
708 	struct nbd_device *nbd = bdev->bd_disk->private_data;
709 	int error;
710 
711 	if (!capable(CAP_SYS_ADMIN))
712 		return -EPERM;
713 
714 	BUG_ON(nbd->magic != NBD_MAGIC);
715 
716 	/* Anyone capable of this syscall can do *real bad* things */
717 	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
718 		nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
719 
720 	mutex_lock(&nbd->tx_lock);
721 	error = __nbd_ioctl(bdev, nbd, cmd, arg);
722 	mutex_unlock(&nbd->tx_lock);
723 
724 	return error;
725 }
726 
727 static const struct block_device_operations nbd_fops =
728 {
729 	.owner =	THIS_MODULE,
730 	.ioctl =	nbd_ioctl,
731 };
732 
733 /*
734  * And here should be modules and kernel interface
735  *  (Just smiley confuses emacs :-)
736  */
737 
738 static int __init nbd_init(void)
739 {
740 	int err = -ENOMEM;
741 	int i;
742 	int part_shift;
743 
744 	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
745 
746 	if (max_part < 0) {
747 		printk(KERN_ERR "nbd: max_part must be >= 0\n");
748 		return -EINVAL;
749 	}
750 
751 	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
752 	if (!nbd_dev)
753 		return -ENOMEM;
754 
755 	part_shift = 0;
756 	if (max_part > 0) {
757 		part_shift = fls(max_part);
758 
759 		/*
760 		 * Adjust max_part according to part_shift as it is exported
761 		 * to user space so that user can know the max number of
762 		 * partition kernel should be able to manage.
763 		 *
764 		 * Note that -1 is required because partition 0 is reserved
765 		 * for the whole disk.
766 		 */
767 		max_part = (1UL << part_shift) - 1;
768 	}
769 
770 	if ((1UL << part_shift) > DISK_MAX_PARTS)
771 		return -EINVAL;
772 
773 	if (nbds_max > 1UL << (MINORBITS - part_shift))
774 		return -EINVAL;
775 
776 	for (i = 0; i < nbds_max; i++) {
777 		struct gendisk *disk = alloc_disk(1 << part_shift);
778 		if (!disk)
779 			goto out;
780 		nbd_dev[i].disk = disk;
781 		/*
782 		 * The new linux 2.5 block layer implementation requires
783 		 * every gendisk to have its very own request_queue struct.
784 		 * These structs are big so we dynamically allocate them.
785 		 */
786 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
787 		if (!disk->queue) {
788 			put_disk(disk);
789 			goto out;
790 		}
791 		/*
792 		 * Tell the block layer that we are not a rotational device
793 		 */
794 		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
795 	}
796 
797 	if (register_blkdev(NBD_MAJOR, "nbd")) {
798 		err = -EIO;
799 		goto out;
800 	}
801 
802 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
803 	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
804 
805 	for (i = 0; i < nbds_max; i++) {
806 		struct gendisk *disk = nbd_dev[i].disk;
807 		nbd_dev[i].file = NULL;
808 		nbd_dev[i].magic = NBD_MAGIC;
809 		nbd_dev[i].flags = 0;
810 		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
811 		spin_lock_init(&nbd_dev[i].queue_lock);
812 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
813 		mutex_init(&nbd_dev[i].tx_lock);
814 		init_waitqueue_head(&nbd_dev[i].active_wq);
815 		init_waitqueue_head(&nbd_dev[i].waiting_wq);
816 		nbd_dev[i].blksize = 1024;
817 		nbd_dev[i].bytesize = 0;
818 		disk->major = NBD_MAJOR;
819 		disk->first_minor = i << part_shift;
820 		disk->fops = &nbd_fops;
821 		disk->private_data = &nbd_dev[i];
822 		sprintf(disk->disk_name, "nbd%d", i);
823 		set_capacity(disk, 0);
824 		add_disk(disk);
825 	}
826 
827 	return 0;
828 out:
829 	while (i--) {
830 		blk_cleanup_queue(nbd_dev[i].disk->queue);
831 		put_disk(nbd_dev[i].disk);
832 	}
833 	kfree(nbd_dev);
834 	return err;
835 }
836 
837 static void __exit nbd_cleanup(void)
838 {
839 	int i;
840 	for (i = 0; i < nbds_max; i++) {
841 		struct gendisk *disk = nbd_dev[i].disk;
842 		nbd_dev[i].magic = 0;
843 		if (disk) {
844 			del_gendisk(disk);
845 			blk_cleanup_queue(disk->queue);
846 			put_disk(disk);
847 		}
848 	}
849 	unregister_blkdev(NBD_MAJOR, "nbd");
850 	kfree(nbd_dev);
851 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
852 }
853 
854 module_init(nbd_init);
855 module_exit(nbd_cleanup);
856 
857 MODULE_DESCRIPTION("Network Block Device");
858 MODULE_LICENSE("GPL");
859 
860 module_param(nbds_max, int, 0444);
861 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
862 module_param(max_part, int, 0444);
863 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
864 #ifndef NDEBUG
865 module_param(debugflags, int, 0644);
866 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
867 #endif
868