Linux Audio

Check our new training course

Loading...
v3.1
  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/system.h>
 38#include <asm/types.h>
 39
 40#include <linux/nbd.h>
 41
 42#define LO_MAGIC 0x68797548
 43
 44#ifdef NDEBUG
 45#define dprintk(flags, fmt...)
 46#else /* NDEBUG */
 47#define dprintk(flags, fmt...) do { \
 48	if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
 49} while (0)
 50#define DBG_IOCTL       0x0004
 51#define DBG_INIT        0x0010
 52#define DBG_EXIT        0x0020
 53#define DBG_BLKDEV      0x0100
 54#define DBG_RX          0x0200
 55#define DBG_TX          0x0400
 56static unsigned int debugflags;
 57#endif /* NDEBUG */
 58
 59static unsigned int nbds_max = 16;
 60static struct nbd_device *nbd_dev;
 61static int max_part;
 62
 63/*
 64 * Use just one lock (or at most 1 per NIC). Two arguments for this:
 65 * 1. Each NIC is essentially a synchronization point for all servers
 66 *    accessed through that NIC so there's no need to have more locks
 67 *    than NICs anyway.
 68 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
 69 *    down each lock to the point where they're actually slower than just
 70 *    a single lock.
 71 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
 72 */
 73static DEFINE_SPINLOCK(nbd_lock);
 74
 75#ifndef NDEBUG
 76static const char *ioctl_cmd_to_ascii(int cmd)
 77{
 78	switch (cmd) {
 79	case NBD_SET_SOCK: return "set-sock";
 80	case NBD_SET_BLKSIZE: return "set-blksize";
 81	case NBD_SET_SIZE: return "set-size";
 82	case NBD_DO_IT: return "do-it";
 83	case NBD_CLEAR_SOCK: return "clear-sock";
 84	case NBD_CLEAR_QUE: return "clear-que";
 85	case NBD_PRINT_DEBUG: return "print-debug";
 86	case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
 87	case NBD_DISCONNECT: return "disconnect";
 88	case BLKROSET: return "set-read-only";
 89	case BLKFLSBUF: return "flush-buffer-cache";
 90	}
 91	return "unknown";
 92}
 93
 94static const char *nbdcmd_to_ascii(int cmd)
 95{
 96	switch (cmd) {
 97	case  NBD_CMD_READ: return "read";
 98	case NBD_CMD_WRITE: return "write";
 99	case  NBD_CMD_DISC: return "disconnect";
100	}
101	return "invalid";
102}
103#endif /* NDEBUG */
104
105static void nbd_end_request(struct request *req)
106{
107	int error = req->errors ? -EIO : 0;
108	struct request_queue *q = req->q;
109	unsigned long flags;
110
111	dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
112			req, error ? "failed" : "done");
113
114	spin_lock_irqsave(q->queue_lock, flags);
115	__blk_end_request_all(req, error);
116	spin_unlock_irqrestore(q->queue_lock, flags);
117}
118
119static void sock_shutdown(struct nbd_device *lo, int lock)
120{
121	/* Forcibly shutdown the socket causing all listeners
122	 * to error
123	 *
124	 * FIXME: This code is duplicated from sys_shutdown, but
125	 * there should be a more generic interface rather than
126	 * calling socket ops directly here */
127	if (lock)
128		mutex_lock(&lo->tx_lock);
129	if (lo->sock) {
130		printk(KERN_WARNING "%s: shutting down socket\n",
131			lo->disk->disk_name);
132		kernel_sock_shutdown(lo->sock, SHUT_RDWR);
133		lo->sock = NULL;
134	}
135	if (lock)
136		mutex_unlock(&lo->tx_lock);
137}
138
139static void nbd_xmit_timeout(unsigned long arg)
140{
141	struct task_struct *task = (struct task_struct *)arg;
142
143	printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
144		task->comm, task->pid);
145	force_sig(SIGKILL, task);
146}
147
148/*
149 *  Send or receive packet.
150 */
151static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
152		int msg_flags)
153{
154	struct socket *sock = lo->sock;
155	int result;
156	struct msghdr msg;
157	struct kvec iov;
158	sigset_t blocked, oldset;
159
160	if (unlikely(!sock)) {
161		printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
162		       lo->disk->disk_name, (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	do {
172		sock->sk->sk_allocation = GFP_NOIO;
173		iov.iov_base = buf;
174		iov.iov_len = size;
175		msg.msg_name = NULL;
176		msg.msg_namelen = 0;
177		msg.msg_control = NULL;
178		msg.msg_controllen = 0;
179		msg.msg_flags = msg_flags | MSG_NOSIGNAL;
180
181		if (send) {
182			struct timer_list ti;
183
184			if (lo->xmit_timeout) {
185				init_timer(&ti);
186				ti.function = nbd_xmit_timeout;
187				ti.data = (unsigned long)current;
188				ti.expires = jiffies + lo->xmit_timeout;
189				add_timer(&ti);
190			}
191			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
192			if (lo->xmit_timeout)
193				del_timer_sync(&ti);
194		} else
195			result = kernel_recvmsg(sock, &msg, &iov, 1, size,
196						msg.msg_flags);
197
198		if (signal_pending(current)) {
199			siginfo_t info;
200			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
201				task_pid_nr(current), current->comm,
202				dequeue_signal_lock(current, &current->blocked, &info));
203			result = -EINTR;
204			sock_shutdown(lo, !send);
205			break;
206		}
207
208		if (result <= 0) {
209			if (result == 0)
210				result = -EPIPE; /* short read */
211			break;
212		}
213		size -= result;
214		buf += result;
215	} while (size > 0);
216
217	sigprocmask(SIG_SETMASK, &oldset, NULL);
218
219	return result;
220}
221
222static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
223		int flags)
224{
225	int result;
226	void *kaddr = kmap(bvec->bv_page);
227	result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
 
228	kunmap(bvec->bv_page);
229	return result;
230}
231
232/* always call with the tx_lock held */
233static int nbd_send_req(struct nbd_device *lo, 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			lo->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(lo, 1, &request, sizeof(request),
251			(nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
252	if (result <= 0) {
253		printk(KERN_ERR "%s: Send control failed (result %d)\n",
254				lo->disk->disk_name, 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					lo->disk->disk_name, req, bvec->bv_len);
271			result = sock_send_bvec(lo, bvec, flags);
272			if (result <= 0) {
273				printk(KERN_ERR "%s: Send data failed (result %d)\n",
274						lo->disk->disk_name, result);
 
275				goto error_out;
276			}
277		}
278	}
279	return 0;
280
281error_out:
282	return -EIO;
283}
284
285static struct request *nbd_find_request(struct nbd_device *lo,
286					struct request *xreq)
287{
288	struct request *req, *tmp;
289	int err;
290
291	err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
292	if (unlikely(err))
293		goto out;
294
295	spin_lock(&lo->queue_lock);
296	list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
297		if (req != xreq)
298			continue;
299		list_del_init(&req->queuelist);
300		spin_unlock(&lo->queue_lock);
301		return req;
302	}
303	spin_unlock(&lo->queue_lock);
304
305	err = -ENOENT;
306
307out:
308	return ERR_PTR(err);
309}
310
311static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
312{
313	int result;
314	void *kaddr = kmap(bvec->bv_page);
315	result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
316			MSG_WAITALL);
317	kunmap(bvec->bv_page);
318	return result;
319}
320
321/* NULL returned = something went wrong, inform userspace */
322static struct request *nbd_read_stat(struct nbd_device *lo)
323{
324	int result;
325	struct nbd_reply reply;
326	struct request *req;
327
328	reply.magic = 0;
329	result = sock_xmit(lo, 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
336	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
337		printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
338				lo->disk->disk_name,
339				(unsigned long)ntohl(reply.magic));
340		result = -EPROTO;
341		goto harderror;
342	}
343
344	req = nbd_find_request(lo, *(struct request **)reply.handle);
345	if (IS_ERR(req)) {
346		result = PTR_ERR(req);
347		if (result != -ENOENT)
348			goto harderror;
349
350		printk(KERN_ERR "%s: Unexpected reply (%p)\n",
351				lo->disk->disk_name, reply.handle);
352		result = -EBADR;
353		goto harderror;
354	}
355
356	if (ntohl(reply.error)) {
357		printk(KERN_ERR "%s: Other side returned error (%d)\n",
358				lo->disk->disk_name, ntohl(reply.error));
359		req->errors++;
360		return req;
361	}
362
363	dprintk(DBG_RX, "%s: request %p: got reply\n",
364			lo->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(lo, bvec);
371			if (result <= 0) {
372				printk(KERN_ERR "%s: Receive data failed (result %d)\n",
373						lo->disk->disk_name, result);
374				req->errors++;
375				return req;
376			}
377			dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
378				lo->disk->disk_name, req, bvec->bv_len);
379		}
380	}
381	return req;
382harderror:
383	lo->harderror = result;
384	return NULL;
385}
386
387static 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
396static struct device_attribute pid_attr = {
397	.attr = { .name = "pid", .mode = S_IRUGO},
398	.show = pid_show,
399};
400
401static int nbd_do_it(struct nbd_device *lo)
402{
403	struct request *req;
404	int ret;
405
406	BUG_ON(lo->magic != LO_MAGIC);
407
408	lo->pid = current->pid;
409	ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
410	if (ret) {
411		printk(KERN_ERR "nbd: sysfs_create_file failed!");
412		lo->pid = 0;
413		return ret;
414	}
415
416	while ((req = nbd_read_stat(lo)) != NULL)
417		nbd_end_request(req);
418
419	sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
420	lo->pid = 0;
421	return 0;
422}
423
424static void nbd_clear_que(struct nbd_device *lo)
425{
426	struct request *req;
427
428	BUG_ON(lo->magic != LO_MAGIC);
429
430	/*
431	 * Because we have set lo->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(lo->sock);
439	BUG_ON(lo->active_req);
440
441	while (!list_empty(&lo->queue_head)) {
442		req = list_entry(lo->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
451static void nbd_handle_req(struct nbd_device *lo, 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 (lo->flags & NBD_READ_ONLY) {
460			printk(KERN_ERR "%s: Write on read-only\n",
461					lo->disk->disk_name);
462			goto error_out;
463		}
464	}
465
466	req->errors = 0;
467
468	mutex_lock(&lo->tx_lock);
469	if (unlikely(!lo->sock)) {
470		mutex_unlock(&lo->tx_lock);
471		printk(KERN_ERR "%s: Attempted send on closed socket\n",
472		       lo->disk->disk_name);
473		goto error_out;
474	}
475
476	lo->active_req = req;
477
478	if (nbd_send_req(lo, req) != 0) {
479		printk(KERN_ERR "%s: Request send failed\n",
480				lo->disk->disk_name);
481		req->errors++;
482		nbd_end_request(req);
483	} else {
484		spin_lock(&lo->queue_lock);
485		list_add(&req->queuelist, &lo->queue_head);
486		spin_unlock(&lo->queue_lock);
487	}
488
489	lo->active_req = NULL;
490	mutex_unlock(&lo->tx_lock);
491	wake_up_all(&lo->active_wq);
492
493	return;
494
495error_out:
496	req->errors++;
497	nbd_end_request(req);
498}
499
500static int nbd_thread(void *data)
501{
502	struct nbd_device *lo = data;
503	struct request *req;
504
505	set_user_nice(current, -20);
506	while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
507		/* wait for something to do */
508		wait_event_interruptible(lo->waiting_wq,
509					 kthread_should_stop() ||
510					 !list_empty(&lo->waiting_queue));
511
512		/* extract request */
513		if (list_empty(&lo->waiting_queue))
514			continue;
515
516		spin_lock_irq(&lo->queue_lock);
517		req = list_entry(lo->waiting_queue.next, struct request,
518				 queuelist);
519		list_del_init(&req->queuelist);
520		spin_unlock_irq(&lo->queue_lock);
521
522		/* handle request */
523		nbd_handle_req(lo, req);
524	}
525	return 0;
526}
527
528/*
529 * We always wait for result of write, for now. It would be nice to make it optional
530 * in future
531 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
532 *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
533 */
534
535static void do_nbd_request(struct request_queue *q)
536{
537	struct request *req;
538	
539	while ((req = blk_fetch_request(q)) != NULL) {
540		struct nbd_device *lo;
541
542		spin_unlock_irq(q->queue_lock);
543
544		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
545				req->rq_disk->disk_name, req, req->cmd_type);
546
547		lo = req->rq_disk->private_data;
548
549		BUG_ON(lo->magic != LO_MAGIC);
550
551		if (unlikely(!lo->sock)) {
552			printk(KERN_ERR "%s: Attempted send on closed socket\n",
553				lo->disk->disk_name);
554			req->errors++;
555			nbd_end_request(req);
556			spin_lock_irq(q->queue_lock);
557			continue;
558		}
559
560		spin_lock_irq(&lo->queue_lock);
561		list_add_tail(&req->queuelist, &lo->waiting_queue);
562		spin_unlock_irq(&lo->queue_lock);
563
564		wake_up(&lo->waiting_wq);
565
566		spin_lock_irq(q->queue_lock);
567	}
568}
569
570/* Must be called with tx_lock held */
571
572static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
573		       unsigned int cmd, unsigned long arg)
574{
575	switch (cmd) {
576	case NBD_DISCONNECT: {
577		struct request sreq;
578
579	        printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
580
581		blk_rq_init(NULL, &sreq);
582		sreq.cmd_type = REQ_TYPE_SPECIAL;
583		nbd_cmd(&sreq) = NBD_CMD_DISC;
584		if (!lo->sock)
585			return -EINVAL;
586		nbd_send_req(lo, &sreq);
587                return 0;
588	}
589 
590	case NBD_CLEAR_SOCK: {
591		struct file *file;
592
593		lo->sock = NULL;
594		file = lo->file;
595		lo->file = NULL;
596		nbd_clear_que(lo);
597		BUG_ON(!list_empty(&lo->queue_head));
 
598		if (file)
599			fput(file);
600		return 0;
601	}
602
603	case NBD_SET_SOCK: {
604		struct file *file;
605		if (lo->file)
606			return -EBUSY;
607		file = fget(arg);
608		if (file) {
609			struct inode *inode = file->f_path.dentry->d_inode;
610			if (S_ISSOCK(inode->i_mode)) {
611				lo->file = file;
612				lo->sock = SOCKET_I(inode);
613				if (max_part > 0)
614					bdev->bd_invalidated = 1;
615				return 0;
616			} else {
617				fput(file);
618			}
619		}
620		return -EINVAL;
621	}
622
623	case NBD_SET_BLKSIZE:
624		lo->blksize = arg;
625		lo->bytesize &= ~(lo->blksize-1);
626		bdev->bd_inode->i_size = lo->bytesize;
627		set_blocksize(bdev, lo->blksize);
628		set_capacity(lo->disk, lo->bytesize >> 9);
629		return 0;
630
631	case NBD_SET_SIZE:
632		lo->bytesize = arg & ~(lo->blksize-1);
633		bdev->bd_inode->i_size = lo->bytesize;
634		set_blocksize(bdev, lo->blksize);
635		set_capacity(lo->disk, lo->bytesize >> 9);
636		return 0;
637
638	case NBD_SET_TIMEOUT:
639		lo->xmit_timeout = arg * HZ;
640		return 0;
641
642	case NBD_SET_SIZE_BLOCKS:
643		lo->bytesize = ((u64) arg) * lo->blksize;
644		bdev->bd_inode->i_size = lo->bytesize;
645		set_blocksize(bdev, lo->blksize);
646		set_capacity(lo->disk, lo->bytesize >> 9);
647		return 0;
648
649	case NBD_DO_IT: {
650		struct task_struct *thread;
651		struct file *file;
652		int error;
653
654		if (lo->pid)
655			return -EBUSY;
656		if (!lo->file)
657			return -EINVAL;
658
659		mutex_unlock(&lo->tx_lock);
660
661		thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
662		if (IS_ERR(thread)) {
663			mutex_lock(&lo->tx_lock);
664			return PTR_ERR(thread);
665		}
666		wake_up_process(thread);
667		error = nbd_do_it(lo);
668		kthread_stop(thread);
669
670		mutex_lock(&lo->tx_lock);
671		if (error)
672			return error;
673		sock_shutdown(lo, 0);
674		file = lo->file;
675		lo->file = NULL;
676		nbd_clear_que(lo);
677		printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
678		if (file)
679			fput(file);
680		lo->bytesize = 0;
681		bdev->bd_inode->i_size = 0;
682		set_capacity(lo->disk, 0);
683		if (max_part > 0)
684			ioctl_by_bdev(bdev, BLKRRPART, 0);
685		return lo->harderror;
686	}
687
688	case NBD_CLEAR_QUE:
689		/*
690		 * This is for compatibility only.  The queue is always cleared
691		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
692		 */
693		BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
694		return 0;
695
696	case NBD_PRINT_DEBUG:
697		printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
698			bdev->bd_disk->disk_name,
699			lo->queue_head.next, lo->queue_head.prev,
700			&lo->queue_head);
701		return 0;
702	}
703	return -ENOTTY;
704}
705
706static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
707		     unsigned int cmd, unsigned long arg)
708{
709	struct nbd_device *lo = bdev->bd_disk->private_data;
710	int error;
711
712	if (!capable(CAP_SYS_ADMIN))
713		return -EPERM;
714
715	BUG_ON(lo->magic != LO_MAGIC);
716
717	/* Anyone capable of this syscall can do *real bad* things */
718	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
719			lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
720
721	mutex_lock(&lo->tx_lock);
722	error = __nbd_ioctl(bdev, lo, cmd, arg);
723	mutex_unlock(&lo->tx_lock);
724
725	return error;
726}
727
728static const struct block_device_operations nbd_fops =
729{
730	.owner =	THIS_MODULE,
731	.ioctl =	nbd_ioctl,
732};
733
734/*
735 * And here should be modules and kernel interface 
736 *  (Just smiley confuses emacs :-)
737 */
738
739static int __init nbd_init(void)
740{
741	int err = -ENOMEM;
742	int i;
743	int part_shift;
744
745	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
746
747	if (max_part < 0) {
748		printk(KERN_CRIT "nbd: max_part must be >= 0\n");
749		return -EINVAL;
750	}
751
752	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
753	if (!nbd_dev)
754		return -ENOMEM;
755
756	part_shift = 0;
757	if (max_part > 0) {
758		part_shift = fls(max_part);
759
760		/*
761		 * Adjust max_part according to part_shift as it is exported
762		 * to user space so that user can know the max number of
763		 * partition kernel should be able to manage.
764		 *
765		 * Note that -1 is required because partition 0 is reserved
766		 * for the whole disk.
767		 */
768		max_part = (1UL << part_shift) - 1;
769	}
770
771	if ((1UL << part_shift) > DISK_MAX_PARTS)
772		return -EINVAL;
773
774	if (nbds_max > 1UL << (MINORBITS - part_shift))
775		return -EINVAL;
776
777	for (i = 0; i < nbds_max; i++) {
778		struct gendisk *disk = alloc_disk(1 << part_shift);
779		if (!disk)
780			goto out;
781		nbd_dev[i].disk = disk;
782		/*
783		 * The new linux 2.5 block layer implementation requires
784		 * every gendisk to have its very own request_queue struct.
785		 * These structs are big so we dynamically allocate them.
786		 */
787		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
788		if (!disk->queue) {
789			put_disk(disk);
790			goto out;
791		}
792		/*
793		 * Tell the block layer that we are not a rotational device
794		 */
795		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
796	}
797
798	if (register_blkdev(NBD_MAJOR, "nbd")) {
799		err = -EIO;
800		goto out;
801	}
802
803	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
804	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
805
806	for (i = 0; i < nbds_max; i++) {
807		struct gendisk *disk = nbd_dev[i].disk;
808		nbd_dev[i].file = NULL;
809		nbd_dev[i].magic = LO_MAGIC;
810		nbd_dev[i].flags = 0;
811		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
812		spin_lock_init(&nbd_dev[i].queue_lock);
813		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
814		mutex_init(&nbd_dev[i].tx_lock);
815		init_waitqueue_head(&nbd_dev[i].active_wq);
816		init_waitqueue_head(&nbd_dev[i].waiting_wq);
817		nbd_dev[i].blksize = 1024;
818		nbd_dev[i].bytesize = 0;
819		disk->major = NBD_MAJOR;
820		disk->first_minor = i << part_shift;
821		disk->fops = &nbd_fops;
822		disk->private_data = &nbd_dev[i];
823		sprintf(disk->disk_name, "nbd%d", i);
824		set_capacity(disk, 0);
825		add_disk(disk);
826	}
827
828	return 0;
829out:
830	while (i--) {
831		blk_cleanup_queue(nbd_dev[i].disk->queue);
832		put_disk(nbd_dev[i].disk);
833	}
834	kfree(nbd_dev);
835	return err;
836}
837
838static void __exit nbd_cleanup(void)
839{
840	int i;
841	for (i = 0; i < nbds_max; i++) {
842		struct gendisk *disk = nbd_dev[i].disk;
843		nbd_dev[i].magic = 0;
844		if (disk) {
845			del_gendisk(disk);
846			blk_cleanup_queue(disk->queue);
847			put_disk(disk);
848		}
849	}
850	unregister_blkdev(NBD_MAJOR, "nbd");
851	kfree(nbd_dev);
852	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
853}
854
855module_init(nbd_init);
856module_exit(nbd_cleanup);
857
858MODULE_DESCRIPTION("Network Block Device");
859MODULE_LICENSE("GPL");
860
861module_param(nbds_max, int, 0444);
862MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
863module_param(max_part, int, 0444);
864MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
865#ifndef NDEBUG
866module_param(debugflags, int, 0644);
867MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
868#endif
v3.5.6
  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
 55static unsigned int debugflags;
 56#endif /* NDEBUG */
 57
 58static unsigned int nbds_max = 16;
 59static struct nbd_device *nbd_dev;
 60static 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 */
 72static DEFINE_SPINLOCK(nbd_lock);
 73
 74#ifndef NDEBUG
 75static 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
 93static 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
104static 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
118static 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
137static 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 */
149static 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
221static 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 */
233static 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
282error_out:
283	return -EIO;
284}
285
286static 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
308out:
309	return ERR_PTR(err);
310}
311
312static 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 */
323static 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;
382harderror:
383	nbd->harderror = result;
384	return NULL;
385}
386
387static 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
396static struct device_attribute pid_attr = {
397	.attr = { .name = "pid", .mode = S_IRUGO},
398	.show = pid_show,
399};
400
401static 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
424static 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	while (!list_empty(&nbd->waiting_queue)) {
450		req = list_entry(nbd->waiting_queue.next, struct request,
451				 queuelist);
452		list_del_init(&req->queuelist);
453		req->errors++;
454		nbd_end_request(req);
455	}
456}
457
458
459static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
460{
461	if (req->cmd_type != REQ_TYPE_FS)
462		goto error_out;
463
464	nbd_cmd(req) = NBD_CMD_READ;
465	if (rq_data_dir(req) == WRITE) {
466		nbd_cmd(req) = NBD_CMD_WRITE;
467		if (nbd->flags & NBD_READ_ONLY) {
468			dev_err(disk_to_dev(nbd->disk),
469				"Write on read-only\n");
470			goto error_out;
471		}
472	}
473
474	req->errors = 0;
475
476	mutex_lock(&nbd->tx_lock);
477	if (unlikely(!nbd->sock)) {
478		mutex_unlock(&nbd->tx_lock);
479		dev_err(disk_to_dev(nbd->disk),
480			"Attempted send on closed socket\n");
481		goto error_out;
482	}
483
484	nbd->active_req = req;
485
486	if (nbd_send_req(nbd, req) != 0) {
487		dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
 
488		req->errors++;
489		nbd_end_request(req);
490	} else {
491		spin_lock(&nbd->queue_lock);
492		list_add(&req->queuelist, &nbd->queue_head);
493		spin_unlock(&nbd->queue_lock);
494	}
495
496	nbd->active_req = NULL;
497	mutex_unlock(&nbd->tx_lock);
498	wake_up_all(&nbd->active_wq);
499
500	return;
501
502error_out:
503	req->errors++;
504	nbd_end_request(req);
505}
506
507static int nbd_thread(void *data)
508{
509	struct nbd_device *nbd = data;
510	struct request *req;
511
512	set_user_nice(current, -20);
513	while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
514		/* wait for something to do */
515		wait_event_interruptible(nbd->waiting_wq,
516					 kthread_should_stop() ||
517					 !list_empty(&nbd->waiting_queue));
518
519		/* extract request */
520		if (list_empty(&nbd->waiting_queue))
521			continue;
522
523		spin_lock_irq(&nbd->queue_lock);
524		req = list_entry(nbd->waiting_queue.next, struct request,
525				 queuelist);
526		list_del_init(&req->queuelist);
527		spin_unlock_irq(&nbd->queue_lock);
528
529		/* handle request */
530		nbd_handle_req(nbd, req);
531	}
532	return 0;
533}
534
535/*
536 * We always wait for result of write, for now. It would be nice to make it optional
537 * in future
538 * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
539 *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
540 */
541
542static void do_nbd_request(struct request_queue *q)
543{
544	struct request *req;
545	
546	while ((req = blk_fetch_request(q)) != NULL) {
547		struct nbd_device *nbd;
548
549		spin_unlock_irq(q->queue_lock);
550
551		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
552				req->rq_disk->disk_name, req, req->cmd_type);
553
554		nbd = req->rq_disk->private_data;
555
556		BUG_ON(nbd->magic != NBD_MAGIC);
557
558		if (unlikely(!nbd->sock)) {
559			dev_err(disk_to_dev(nbd->disk),
560				"Attempted send on closed socket\n");
561			req->errors++;
562			nbd_end_request(req);
563			spin_lock_irq(q->queue_lock);
564			continue;
565		}
566
567		spin_lock_irq(&nbd->queue_lock);
568		list_add_tail(&req->queuelist, &nbd->waiting_queue);
569		spin_unlock_irq(&nbd->queue_lock);
570
571		wake_up(&nbd->waiting_wq);
572
573		spin_lock_irq(q->queue_lock);
574	}
575}
576
577/* Must be called with tx_lock held */
578
579static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
580		       unsigned int cmd, unsigned long arg)
581{
582	switch (cmd) {
583	case NBD_DISCONNECT: {
584		struct request sreq;
585
586		dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
587
588		blk_rq_init(NULL, &sreq);
589		sreq.cmd_type = REQ_TYPE_SPECIAL;
590		nbd_cmd(&sreq) = NBD_CMD_DISC;
591		if (!nbd->sock)
592			return -EINVAL;
593		nbd_send_req(nbd, &sreq);
594                return 0;
595	}
596 
597	case NBD_CLEAR_SOCK: {
598		struct file *file;
599
600		nbd->sock = NULL;
601		file = nbd->file;
602		nbd->file = NULL;
603		nbd_clear_que(nbd);
604		BUG_ON(!list_empty(&nbd->queue_head));
605		BUG_ON(!list_empty(&nbd->waiting_queue));
606		if (file)
607			fput(file);
608		return 0;
609	}
610
611	case NBD_SET_SOCK: {
612		struct file *file;
613		if (nbd->file)
614			return -EBUSY;
615		file = fget(arg);
616		if (file) {
617			struct inode *inode = file->f_path.dentry->d_inode;
618			if (S_ISSOCK(inode->i_mode)) {
619				nbd->file = file;
620				nbd->sock = SOCKET_I(inode);
621				if (max_part > 0)
622					bdev->bd_invalidated = 1;
623				return 0;
624			} else {
625				fput(file);
626			}
627		}
628		return -EINVAL;
629	}
630
631	case NBD_SET_BLKSIZE:
632		nbd->blksize = arg;
633		nbd->bytesize &= ~(nbd->blksize-1);
634		bdev->bd_inode->i_size = nbd->bytesize;
635		set_blocksize(bdev, nbd->blksize);
636		set_capacity(nbd->disk, nbd->bytesize >> 9);
637		return 0;
638
639	case NBD_SET_SIZE:
640		nbd->bytesize = arg & ~(nbd->blksize-1);
641		bdev->bd_inode->i_size = nbd->bytesize;
642		set_blocksize(bdev, nbd->blksize);
643		set_capacity(nbd->disk, nbd->bytesize >> 9);
644		return 0;
645
646	case NBD_SET_TIMEOUT:
647		nbd->xmit_timeout = arg * HZ;
648		return 0;
649
650	case NBD_SET_SIZE_BLOCKS:
651		nbd->bytesize = ((u64) arg) * nbd->blksize;
652		bdev->bd_inode->i_size = nbd->bytesize;
653		set_blocksize(bdev, nbd->blksize);
654		set_capacity(nbd->disk, nbd->bytesize >> 9);
655		return 0;
656
657	case NBD_DO_IT: {
658		struct task_struct *thread;
659		struct file *file;
660		int error;
661
662		if (nbd->pid)
663			return -EBUSY;
664		if (!nbd->file)
665			return -EINVAL;
666
667		mutex_unlock(&nbd->tx_lock);
668
669		thread = kthread_create(nbd_thread, nbd, nbd->disk->disk_name);
670		if (IS_ERR(thread)) {
671			mutex_lock(&nbd->tx_lock);
672			return PTR_ERR(thread);
673		}
674		wake_up_process(thread);
675		error = nbd_do_it(nbd);
676		kthread_stop(thread);
677
678		mutex_lock(&nbd->tx_lock);
679		if (error)
680			return error;
681		sock_shutdown(nbd, 0);
682		file = nbd->file;
683		nbd->file = NULL;
684		nbd_clear_que(nbd);
685		dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
686		if (file)
687			fput(file);
688		nbd->bytesize = 0;
689		bdev->bd_inode->i_size = 0;
690		set_capacity(nbd->disk, 0);
691		if (max_part > 0)
692			ioctl_by_bdev(bdev, BLKRRPART, 0);
693		return nbd->harderror;
694	}
695
696	case NBD_CLEAR_QUE:
697		/*
698		 * This is for compatibility only.  The queue is always cleared
699		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
700		 */
701		BUG_ON(!nbd->sock && !list_empty(&nbd->queue_head));
702		return 0;
703
704	case NBD_PRINT_DEBUG:
705		dev_info(disk_to_dev(nbd->disk),
706			"next = %p, prev = %p, head = %p\n",
707			nbd->queue_head.next, nbd->queue_head.prev,
708			&nbd->queue_head);
709		return 0;
710	}
711	return -ENOTTY;
712}
713
714static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
715		     unsigned int cmd, unsigned long arg)
716{
717	struct nbd_device *nbd = bdev->bd_disk->private_data;
718	int error;
719
720	if (!capable(CAP_SYS_ADMIN))
721		return -EPERM;
722
723	BUG_ON(nbd->magic != NBD_MAGIC);
724
725	/* Anyone capable of this syscall can do *real bad* things */
726	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
727		nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
728
729	mutex_lock(&nbd->tx_lock);
730	error = __nbd_ioctl(bdev, nbd, cmd, arg);
731	mutex_unlock(&nbd->tx_lock);
732
733	return error;
734}
735
736static const struct block_device_operations nbd_fops =
737{
738	.owner =	THIS_MODULE,
739	.ioctl =	nbd_ioctl,
740};
741
742/*
743 * And here should be modules and kernel interface 
744 *  (Just smiley confuses emacs :-)
745 */
746
747static int __init nbd_init(void)
748{
749	int err = -ENOMEM;
750	int i;
751	int part_shift;
752
753	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
754
755	if (max_part < 0) {
756		printk(KERN_ERR "nbd: max_part must be >= 0\n");
757		return -EINVAL;
758	}
759
760	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
761	if (!nbd_dev)
762		return -ENOMEM;
763
764	part_shift = 0;
765	if (max_part > 0) {
766		part_shift = fls(max_part);
767
768		/*
769		 * Adjust max_part according to part_shift as it is exported
770		 * to user space so that user can know the max number of
771		 * partition kernel should be able to manage.
772		 *
773		 * Note that -1 is required because partition 0 is reserved
774		 * for the whole disk.
775		 */
776		max_part = (1UL << part_shift) - 1;
777	}
778
779	if ((1UL << part_shift) > DISK_MAX_PARTS)
780		return -EINVAL;
781
782	if (nbds_max > 1UL << (MINORBITS - part_shift))
783		return -EINVAL;
784
785	for (i = 0; i < nbds_max; i++) {
786		struct gendisk *disk = alloc_disk(1 << part_shift);
787		if (!disk)
788			goto out;
789		nbd_dev[i].disk = disk;
790		/*
791		 * The new linux 2.5 block layer implementation requires
792		 * every gendisk to have its very own request_queue struct.
793		 * These structs are big so we dynamically allocate them.
794		 */
795		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
796		if (!disk->queue) {
797			put_disk(disk);
798			goto out;
799		}
800		/*
801		 * Tell the block layer that we are not a rotational device
802		 */
803		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
804	}
805
806	if (register_blkdev(NBD_MAJOR, "nbd")) {
807		err = -EIO;
808		goto out;
809	}
810
811	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
812	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
813
814	for (i = 0; i < nbds_max; i++) {
815		struct gendisk *disk = nbd_dev[i].disk;
816		nbd_dev[i].file = NULL;
817		nbd_dev[i].magic = NBD_MAGIC;
818		nbd_dev[i].flags = 0;
819		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
820		spin_lock_init(&nbd_dev[i].queue_lock);
821		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
822		mutex_init(&nbd_dev[i].tx_lock);
823		init_waitqueue_head(&nbd_dev[i].active_wq);
824		init_waitqueue_head(&nbd_dev[i].waiting_wq);
825		nbd_dev[i].blksize = 1024;
826		nbd_dev[i].bytesize = 0;
827		disk->major = NBD_MAJOR;
828		disk->first_minor = i << part_shift;
829		disk->fops = &nbd_fops;
830		disk->private_data = &nbd_dev[i];
831		sprintf(disk->disk_name, "nbd%d", i);
832		set_capacity(disk, 0);
833		add_disk(disk);
834	}
835
836	return 0;
837out:
838	while (i--) {
839		blk_cleanup_queue(nbd_dev[i].disk->queue);
840		put_disk(nbd_dev[i].disk);
841	}
842	kfree(nbd_dev);
843	return err;
844}
845
846static void __exit nbd_cleanup(void)
847{
848	int i;
849	for (i = 0; i < nbds_max; i++) {
850		struct gendisk *disk = nbd_dev[i].disk;
851		nbd_dev[i].magic = 0;
852		if (disk) {
853			del_gendisk(disk);
854			blk_cleanup_queue(disk->queue);
855			put_disk(disk);
856		}
857	}
858	unregister_blkdev(NBD_MAJOR, "nbd");
859	kfree(nbd_dev);
860	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
861}
862
863module_init(nbd_init);
864module_exit(nbd_cleanup);
865
866MODULE_DESCRIPTION("Network Block Device");
867MODULE_LICENSE("GPL");
868
869module_param(nbds_max, int, 0444);
870MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
871module_param(max_part, int, 0444);
872MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
873#ifndef NDEBUG
874module_param(debugflags, int, 0644);
875MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
876#endif