xref: /openbmc/linux/drivers/block/nbd.c (revision 4f3db074)
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 #include <linux/types.h>
36 
37 #include <asm/uaccess.h>
38 #include <asm/types.h>
39 
40 #include <linux/nbd.h>
41 
42 struct nbd_device {
43 	int flags;
44 	int harderror;		/* Code of hard error			*/
45 	struct socket * sock;	/* If == NULL, device is not ready, yet	*/
46 	int magic;
47 
48 	spinlock_t queue_lock;
49 	struct list_head queue_head;	/* Requests waiting result */
50 	struct request *active_req;
51 	wait_queue_head_t active_wq;
52 	struct list_head waiting_queue;	/* Requests to be sent */
53 	wait_queue_head_t waiting_wq;
54 
55 	struct mutex tx_lock;
56 	struct gendisk *disk;
57 	int blksize;
58 	loff_t bytesize;
59 	pid_t pid; /* pid of nbd-client, if attached */
60 	int xmit_timeout;
61 	int disconnect; /* a disconnect has been requested by user */
62 };
63 
64 #define NBD_MAGIC 0x68797548
65 
66 static unsigned int nbds_max = 16;
67 static struct nbd_device *nbd_dev;
68 static int max_part;
69 
70 /*
71  * Use just one lock (or at most 1 per NIC). Two arguments for this:
72  * 1. Each NIC is essentially a synchronization point for all servers
73  *    accessed through that NIC so there's no need to have more locks
74  *    than NICs anyway.
75  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
76  *    down each lock to the point where they're actually slower than just
77  *    a single lock.
78  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
79  */
80 static DEFINE_SPINLOCK(nbd_lock);
81 
82 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
83 {
84 	return disk_to_dev(nbd->disk);
85 }
86 
87 static const char *nbdcmd_to_ascii(int cmd)
88 {
89 	switch (cmd) {
90 	case  NBD_CMD_READ: return "read";
91 	case NBD_CMD_WRITE: return "write";
92 	case  NBD_CMD_DISC: return "disconnect";
93 	case NBD_CMD_FLUSH: return "flush";
94 	case  NBD_CMD_TRIM: return "trim/discard";
95 	}
96 	return "invalid";
97 }
98 
99 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
100 {
101 	int error = req->errors ? -EIO : 0;
102 	struct request_queue *q = req->q;
103 	unsigned long flags;
104 
105 	dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
106 		error ? "failed" : "done");
107 
108 	spin_lock_irqsave(q->queue_lock, flags);
109 	__blk_end_request_all(req, error);
110 	spin_unlock_irqrestore(q->queue_lock, flags);
111 }
112 
113 /*
114  * Forcibly shutdown the socket causing all listeners to error
115  */
116 static void sock_shutdown(struct nbd_device *nbd, int lock)
117 {
118 	if (lock)
119 		mutex_lock(&nbd->tx_lock);
120 	if (nbd->sock) {
121 		dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
122 		kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
123 		nbd->sock = NULL;
124 	}
125 	if (lock)
126 		mutex_unlock(&nbd->tx_lock);
127 }
128 
129 static void nbd_xmit_timeout(unsigned long arg)
130 {
131 	struct task_struct *task = (struct task_struct *)arg;
132 
133 	printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
134 		task->comm, task->pid);
135 	force_sig(SIGKILL, task);
136 }
137 
138 /*
139  *  Send or receive packet.
140  */
141 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
142 		int msg_flags)
143 {
144 	struct socket *sock = nbd->sock;
145 	int result;
146 	struct msghdr msg;
147 	struct kvec iov;
148 	sigset_t blocked, oldset;
149 	unsigned long pflags = current->flags;
150 
151 	if (unlikely(!sock)) {
152 		dev_err(disk_to_dev(nbd->disk),
153 			"Attempted %s on closed socket in sock_xmit\n",
154 			(send ? "send" : "recv"));
155 		return -EINVAL;
156 	}
157 
158 	/* Allow interception of SIGKILL only
159 	 * Don't allow other signals to interrupt the transmission */
160 	siginitsetinv(&blocked, sigmask(SIGKILL));
161 	sigprocmask(SIG_SETMASK, &blocked, &oldset);
162 
163 	current->flags |= PF_MEMALLOC;
164 	do {
165 		sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
166 		iov.iov_base = buf;
167 		iov.iov_len = size;
168 		msg.msg_name = NULL;
169 		msg.msg_namelen = 0;
170 		msg.msg_control = NULL;
171 		msg.msg_controllen = 0;
172 		msg.msg_flags = msg_flags | MSG_NOSIGNAL;
173 
174 		if (send) {
175 			struct timer_list ti;
176 
177 			if (nbd->xmit_timeout) {
178 				init_timer(&ti);
179 				ti.function = nbd_xmit_timeout;
180 				ti.data = (unsigned long)current;
181 				ti.expires = jiffies + nbd->xmit_timeout;
182 				add_timer(&ti);
183 			}
184 			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
185 			if (nbd->xmit_timeout)
186 				del_timer_sync(&ti);
187 		} else
188 			result = kernel_recvmsg(sock, &msg, &iov, 1, size,
189 						msg.msg_flags);
190 
191 		if (signal_pending(current)) {
192 			siginfo_t info;
193 			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
194 				task_pid_nr(current), current->comm,
195 				dequeue_signal_lock(current, &current->blocked, &info));
196 			result = -EINTR;
197 			sock_shutdown(nbd, !send);
198 			break;
199 		}
200 
201 		if (result <= 0) {
202 			if (result == 0)
203 				result = -EPIPE; /* short read */
204 			break;
205 		}
206 		size -= result;
207 		buf += result;
208 	} while (size > 0);
209 
210 	sigprocmask(SIG_SETMASK, &oldset, NULL);
211 	tsk_restore_flags(current, pflags, PF_MEMALLOC);
212 
213 	return result;
214 }
215 
216 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
217 		int flags)
218 {
219 	int result;
220 	void *kaddr = kmap(bvec->bv_page);
221 	result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
222 			   bvec->bv_len, flags);
223 	kunmap(bvec->bv_page);
224 	return result;
225 }
226 
227 /* always call with the tx_lock held */
228 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
229 {
230 	int result, flags;
231 	struct nbd_request request;
232 	unsigned long size = blk_rq_bytes(req);
233 
234 	memset(&request, 0, sizeof(request));
235 	request.magic = htonl(NBD_REQUEST_MAGIC);
236 	request.type = htonl(nbd_cmd(req));
237 
238 	if (nbd_cmd(req) != NBD_CMD_FLUSH && nbd_cmd(req) != NBD_CMD_DISC) {
239 		request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
240 		request.len = htonl(size);
241 	}
242 	memcpy(request.handle, &req, sizeof(req));
243 
244 	dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
245 		req, nbdcmd_to_ascii(nbd_cmd(req)),
246 		(unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
247 	result = sock_xmit(nbd, 1, &request, sizeof(request),
248 			(nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
249 	if (result <= 0) {
250 		dev_err(disk_to_dev(nbd->disk),
251 			"Send control failed (result %d)\n", result);
252 		return -EIO;
253 	}
254 
255 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
256 		struct req_iterator iter;
257 		struct bio_vec bvec;
258 		/*
259 		 * we are really probing at internals to determine
260 		 * whether to set MSG_MORE or not...
261 		 */
262 		rq_for_each_segment(bvec, req, iter) {
263 			flags = 0;
264 			if (!rq_iter_last(bvec, iter))
265 				flags = MSG_MORE;
266 			dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
267 				req, bvec.bv_len);
268 			result = sock_send_bvec(nbd, &bvec, flags);
269 			if (result <= 0) {
270 				dev_err(disk_to_dev(nbd->disk),
271 					"Send data failed (result %d)\n",
272 					result);
273 				return -EIO;
274 			}
275 		}
276 	}
277 	return 0;
278 }
279 
280 static struct request *nbd_find_request(struct nbd_device *nbd,
281 					struct request *xreq)
282 {
283 	struct request *req, *tmp;
284 	int err;
285 
286 	err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
287 	if (unlikely(err))
288 		return ERR_PTR(err);
289 
290 	spin_lock(&nbd->queue_lock);
291 	list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
292 		if (req != xreq)
293 			continue;
294 		list_del_init(&req->queuelist);
295 		spin_unlock(&nbd->queue_lock);
296 		return req;
297 	}
298 	spin_unlock(&nbd->queue_lock);
299 
300 	return ERR_PTR(-ENOENT);
301 }
302 
303 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
304 {
305 	int result;
306 	void *kaddr = kmap(bvec->bv_page);
307 	result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
308 			MSG_WAITALL);
309 	kunmap(bvec->bv_page);
310 	return result;
311 }
312 
313 /* NULL returned = something went wrong, inform userspace */
314 static struct request *nbd_read_stat(struct nbd_device *nbd)
315 {
316 	int result;
317 	struct nbd_reply reply;
318 	struct request *req;
319 
320 	reply.magic = 0;
321 	result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
322 	if (result <= 0) {
323 		dev_err(disk_to_dev(nbd->disk),
324 			"Receive control failed (result %d)\n", result);
325 		goto harderror;
326 	}
327 
328 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
329 		dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
330 				(unsigned long)ntohl(reply.magic));
331 		result = -EPROTO;
332 		goto harderror;
333 	}
334 
335 	req = nbd_find_request(nbd, *(struct request **)reply.handle);
336 	if (IS_ERR(req)) {
337 		result = PTR_ERR(req);
338 		if (result != -ENOENT)
339 			goto harderror;
340 
341 		dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
342 			reply.handle);
343 		result = -EBADR;
344 		goto harderror;
345 	}
346 
347 	if (ntohl(reply.error)) {
348 		dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
349 			ntohl(reply.error));
350 		req->errors++;
351 		return req;
352 	}
353 
354 	dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
355 	if (nbd_cmd(req) == NBD_CMD_READ) {
356 		struct req_iterator iter;
357 		struct bio_vec bvec;
358 
359 		rq_for_each_segment(bvec, req, iter) {
360 			result = sock_recv_bvec(nbd, &bvec);
361 			if (result <= 0) {
362 				dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
363 					result);
364 				req->errors++;
365 				return req;
366 			}
367 			dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
368 				req, bvec.bv_len);
369 		}
370 	}
371 	return req;
372 harderror:
373 	nbd->harderror = result;
374 	return NULL;
375 }
376 
377 static ssize_t pid_show(struct device *dev,
378 			struct device_attribute *attr, char *buf)
379 {
380 	struct gendisk *disk = dev_to_disk(dev);
381 
382 	return sprintf(buf, "%ld\n",
383 		(long) ((struct nbd_device *)disk->private_data)->pid);
384 }
385 
386 static struct device_attribute pid_attr = {
387 	.attr = { .name = "pid", .mode = S_IRUGO},
388 	.show = pid_show,
389 };
390 
391 static int nbd_do_it(struct nbd_device *nbd)
392 {
393 	struct request *req;
394 	int ret;
395 
396 	BUG_ON(nbd->magic != NBD_MAGIC);
397 
398 	sk_set_memalloc(nbd->sock->sk);
399 	nbd->pid = task_pid_nr(current);
400 	ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
401 	if (ret) {
402 		dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
403 		nbd->pid = 0;
404 		return ret;
405 	}
406 
407 	while ((req = nbd_read_stat(nbd)) != NULL)
408 		nbd_end_request(nbd, req);
409 
410 	device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
411 	nbd->pid = 0;
412 	return 0;
413 }
414 
415 static void nbd_clear_que(struct nbd_device *nbd)
416 {
417 	struct request *req;
418 
419 	BUG_ON(nbd->magic != NBD_MAGIC);
420 
421 	/*
422 	 * Because we have set nbd->sock to NULL under the tx_lock, all
423 	 * modifications to the list must have completed by now.  For
424 	 * the same reason, the active_req must be NULL.
425 	 *
426 	 * As a consequence, we don't need to take the spin lock while
427 	 * purging the list here.
428 	 */
429 	BUG_ON(nbd->sock);
430 	BUG_ON(nbd->active_req);
431 
432 	while (!list_empty(&nbd->queue_head)) {
433 		req = list_entry(nbd->queue_head.next, struct request,
434 				 queuelist);
435 		list_del_init(&req->queuelist);
436 		req->errors++;
437 		nbd_end_request(nbd, req);
438 	}
439 
440 	while (!list_empty(&nbd->waiting_queue)) {
441 		req = list_entry(nbd->waiting_queue.next, struct request,
442 				 queuelist);
443 		list_del_init(&req->queuelist);
444 		req->errors++;
445 		nbd_end_request(nbd, req);
446 	}
447 }
448 
449 
450 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
451 {
452 	if (req->cmd_type != REQ_TYPE_FS)
453 		goto error_out;
454 
455 	nbd_cmd(req) = NBD_CMD_READ;
456 	if (rq_data_dir(req) == WRITE) {
457 		if ((req->cmd_flags & REQ_DISCARD)) {
458 			WARN_ON(!(nbd->flags & NBD_FLAG_SEND_TRIM));
459 			nbd_cmd(req) = NBD_CMD_TRIM;
460 		} else
461 			nbd_cmd(req) = NBD_CMD_WRITE;
462 		if (nbd->flags & NBD_FLAG_READ_ONLY) {
463 			dev_err(disk_to_dev(nbd->disk),
464 				"Write on read-only\n");
465 			goto error_out;
466 		}
467 	}
468 
469 	if (req->cmd_flags & REQ_FLUSH) {
470 		BUG_ON(unlikely(blk_rq_sectors(req)));
471 		nbd_cmd(req) = NBD_CMD_FLUSH;
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(nbd, req);
490 	} else {
491 		spin_lock(&nbd->queue_lock);
492 		list_add_tail(&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 
502 error_out:
503 	req->errors++;
504 	nbd_end_request(nbd, req);
505 }
506 
507 static int nbd_thread(void *data)
508 {
509 	struct nbd_device *nbd = data;
510 	struct request *req;
511 
512 	set_user_nice(current, MIN_NICE);
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 
542 static void do_nbd_request(struct request_queue *q)
543 		__releases(q->queue_lock) __acquires(q->queue_lock)
544 {
545 	struct request *req;
546 
547 	while ((req = blk_fetch_request(q)) != NULL) {
548 		struct nbd_device *nbd;
549 
550 		spin_unlock_irq(q->queue_lock);
551 
552 		nbd = req->rq_disk->private_data;
553 
554 		BUG_ON(nbd->magic != NBD_MAGIC);
555 
556 		dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
557 			req, req->cmd_type);
558 
559 		if (unlikely(!nbd->sock)) {
560 			dev_err(disk_to_dev(nbd->disk),
561 				"Attempted send on closed socket\n");
562 			req->errors++;
563 			nbd_end_request(nbd, req);
564 			spin_lock_irq(q->queue_lock);
565 			continue;
566 		}
567 
568 		spin_lock_irq(&nbd->queue_lock);
569 		list_add_tail(&req->queuelist, &nbd->waiting_queue);
570 		spin_unlock_irq(&nbd->queue_lock);
571 
572 		wake_up(&nbd->waiting_wq);
573 
574 		spin_lock_irq(q->queue_lock);
575 	}
576 }
577 
578 /* Must be called with tx_lock held */
579 
580 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
581 		       unsigned int cmd, unsigned long arg)
582 {
583 	switch (cmd) {
584 	case NBD_DISCONNECT: {
585 		struct request sreq;
586 
587 		dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
588 		if (!nbd->sock)
589 			return -EINVAL;
590 
591 		mutex_unlock(&nbd->tx_lock);
592 		fsync_bdev(bdev);
593 		mutex_lock(&nbd->tx_lock);
594 		blk_rq_init(NULL, &sreq);
595 		sreq.cmd_type = REQ_TYPE_SPECIAL;
596 		nbd_cmd(&sreq) = NBD_CMD_DISC;
597 
598 		/* Check again after getting mutex back.  */
599 		if (!nbd->sock)
600 			return -EINVAL;
601 
602 		nbd->disconnect = 1;
603 
604 		nbd_send_req(nbd, &sreq);
605 		return 0;
606 	}
607 
608 	case NBD_CLEAR_SOCK: {
609 		struct socket *sock = nbd->sock;
610 		nbd->sock = NULL;
611 		nbd_clear_que(nbd);
612 		BUG_ON(!list_empty(&nbd->queue_head));
613 		BUG_ON(!list_empty(&nbd->waiting_queue));
614 		kill_bdev(bdev);
615 		if (sock)
616 			sockfd_put(sock);
617 		return 0;
618 	}
619 
620 	case NBD_SET_SOCK: {
621 		struct socket *sock;
622 		int err;
623 		if (nbd->sock)
624 			return -EBUSY;
625 		sock = sockfd_lookup(arg, &err);
626 		if (sock) {
627 			nbd->sock = sock;
628 			if (max_part > 0)
629 				bdev->bd_invalidated = 1;
630 			nbd->disconnect = 0; /* we're connected now */
631 			return 0;
632 		}
633 		return -EINVAL;
634 	}
635 
636 	case NBD_SET_BLKSIZE:
637 		nbd->blksize = arg;
638 		nbd->bytesize &= ~(nbd->blksize-1);
639 		bdev->bd_inode->i_size = nbd->bytesize;
640 		set_blocksize(bdev, nbd->blksize);
641 		set_capacity(nbd->disk, nbd->bytesize >> 9);
642 		return 0;
643 
644 	case NBD_SET_SIZE:
645 		nbd->bytesize = arg & ~(nbd->blksize-1);
646 		bdev->bd_inode->i_size = nbd->bytesize;
647 		set_blocksize(bdev, nbd->blksize);
648 		set_capacity(nbd->disk, nbd->bytesize >> 9);
649 		return 0;
650 
651 	case NBD_SET_TIMEOUT:
652 		nbd->xmit_timeout = arg * HZ;
653 		return 0;
654 
655 	case NBD_SET_FLAGS:
656 		nbd->flags = arg;
657 		return 0;
658 
659 	case NBD_SET_SIZE_BLOCKS:
660 		nbd->bytesize = ((u64) arg) * nbd->blksize;
661 		bdev->bd_inode->i_size = nbd->bytesize;
662 		set_blocksize(bdev, nbd->blksize);
663 		set_capacity(nbd->disk, nbd->bytesize >> 9);
664 		return 0;
665 
666 	case NBD_DO_IT: {
667 		struct task_struct *thread;
668 		struct socket *sock;
669 		int error;
670 
671 		if (nbd->pid)
672 			return -EBUSY;
673 		if (!nbd->sock)
674 			return -EINVAL;
675 
676 		mutex_unlock(&nbd->tx_lock);
677 
678 		if (nbd->flags & NBD_FLAG_READ_ONLY)
679 			set_device_ro(bdev, true);
680 		if (nbd->flags & NBD_FLAG_SEND_TRIM)
681 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
682 				nbd->disk->queue);
683 		if (nbd->flags & NBD_FLAG_SEND_FLUSH)
684 			blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
685 		else
686 			blk_queue_flush(nbd->disk->queue, 0);
687 
688 		thread = kthread_run(nbd_thread, nbd, "%s",
689 				     nbd->disk->disk_name);
690 		if (IS_ERR(thread)) {
691 			mutex_lock(&nbd->tx_lock);
692 			return PTR_ERR(thread);
693 		}
694 
695 		error = nbd_do_it(nbd);
696 		kthread_stop(thread);
697 
698 		mutex_lock(&nbd->tx_lock);
699 		if (error)
700 			return error;
701 		sock_shutdown(nbd, 0);
702 		sock = nbd->sock;
703 		nbd->sock = NULL;
704 		nbd_clear_que(nbd);
705 		dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
706 		kill_bdev(bdev);
707 		queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
708 		set_device_ro(bdev, false);
709 		if (sock)
710 			sockfd_put(sock);
711 		nbd->flags = 0;
712 		nbd->bytesize = 0;
713 		bdev->bd_inode->i_size = 0;
714 		set_capacity(nbd->disk, 0);
715 		if (max_part > 0)
716 			ioctl_by_bdev(bdev, BLKRRPART, 0);
717 		if (nbd->disconnect) /* user requested, ignore socket errors */
718 			return 0;
719 		return nbd->harderror;
720 	}
721 
722 	case NBD_CLEAR_QUE:
723 		/*
724 		 * This is for compatibility only.  The queue is always cleared
725 		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
726 		 */
727 		return 0;
728 
729 	case NBD_PRINT_DEBUG:
730 		dev_info(disk_to_dev(nbd->disk),
731 			"next = %p, prev = %p, head = %p\n",
732 			nbd->queue_head.next, nbd->queue_head.prev,
733 			&nbd->queue_head);
734 		return 0;
735 	}
736 	return -ENOTTY;
737 }
738 
739 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
740 		     unsigned int cmd, unsigned long arg)
741 {
742 	struct nbd_device *nbd = bdev->bd_disk->private_data;
743 	int error;
744 
745 	if (!capable(CAP_SYS_ADMIN))
746 		return -EPERM;
747 
748 	BUG_ON(nbd->magic != NBD_MAGIC);
749 
750 	mutex_lock(&nbd->tx_lock);
751 	error = __nbd_ioctl(bdev, nbd, cmd, arg);
752 	mutex_unlock(&nbd->tx_lock);
753 
754 	return error;
755 }
756 
757 static const struct block_device_operations nbd_fops =
758 {
759 	.owner =	THIS_MODULE,
760 	.ioctl =	nbd_ioctl,
761 };
762 
763 /*
764  * And here should be modules and kernel interface
765  *  (Just smiley confuses emacs :-)
766  */
767 
768 static int __init nbd_init(void)
769 {
770 	int err = -ENOMEM;
771 	int i;
772 	int part_shift;
773 
774 	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
775 
776 	if (max_part < 0) {
777 		printk(KERN_ERR "nbd: max_part must be >= 0\n");
778 		return -EINVAL;
779 	}
780 
781 	part_shift = 0;
782 	if (max_part > 0) {
783 		part_shift = fls(max_part);
784 
785 		/*
786 		 * Adjust max_part according to part_shift as it is exported
787 		 * to user space so that user can know the max number of
788 		 * partition kernel should be able to manage.
789 		 *
790 		 * Note that -1 is required because partition 0 is reserved
791 		 * for the whole disk.
792 		 */
793 		max_part = (1UL << part_shift) - 1;
794 	}
795 
796 	if ((1UL << part_shift) > DISK_MAX_PARTS)
797 		return -EINVAL;
798 
799 	if (nbds_max > 1UL << (MINORBITS - part_shift))
800 		return -EINVAL;
801 
802 	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
803 	if (!nbd_dev)
804 		return -ENOMEM;
805 
806 	for (i = 0; i < nbds_max; i++) {
807 		struct gendisk *disk = alloc_disk(1 << part_shift);
808 		if (!disk)
809 			goto out;
810 		nbd_dev[i].disk = disk;
811 		/*
812 		 * The new linux 2.5 block layer implementation requires
813 		 * every gendisk to have its very own request_queue struct.
814 		 * These structs are big so we dynamically allocate them.
815 		 */
816 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
817 		if (!disk->queue) {
818 			put_disk(disk);
819 			goto out;
820 		}
821 		/*
822 		 * Tell the block layer that we are not a rotational device
823 		 */
824 		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
825 		queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
826 		disk->queue->limits.discard_granularity = 512;
827 		disk->queue->limits.max_discard_sectors = UINT_MAX;
828 		disk->queue->limits.discard_zeroes_data = 0;
829 		blk_queue_max_hw_sectors(disk->queue, 65536);
830 		disk->queue->limits.max_sectors = 256;
831 	}
832 
833 	if (register_blkdev(NBD_MAJOR, "nbd")) {
834 		err = -EIO;
835 		goto out;
836 	}
837 
838 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
839 
840 	for (i = 0; i < nbds_max; i++) {
841 		struct gendisk *disk = nbd_dev[i].disk;
842 		nbd_dev[i].magic = NBD_MAGIC;
843 		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
844 		spin_lock_init(&nbd_dev[i].queue_lock);
845 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
846 		mutex_init(&nbd_dev[i].tx_lock);
847 		init_waitqueue_head(&nbd_dev[i].active_wq);
848 		init_waitqueue_head(&nbd_dev[i].waiting_wq);
849 		nbd_dev[i].blksize = 1024;
850 		nbd_dev[i].bytesize = 0;
851 		disk->major = NBD_MAJOR;
852 		disk->first_minor = i << part_shift;
853 		disk->fops = &nbd_fops;
854 		disk->private_data = &nbd_dev[i];
855 		sprintf(disk->disk_name, "nbd%d", i);
856 		set_capacity(disk, 0);
857 		add_disk(disk);
858 	}
859 
860 	return 0;
861 out:
862 	while (i--) {
863 		blk_cleanup_queue(nbd_dev[i].disk->queue);
864 		put_disk(nbd_dev[i].disk);
865 	}
866 	kfree(nbd_dev);
867 	return err;
868 }
869 
870 static void __exit nbd_cleanup(void)
871 {
872 	int i;
873 	for (i = 0; i < nbds_max; i++) {
874 		struct gendisk *disk = nbd_dev[i].disk;
875 		nbd_dev[i].magic = 0;
876 		if (disk) {
877 			del_gendisk(disk);
878 			blk_cleanup_queue(disk->queue);
879 			put_disk(disk);
880 		}
881 	}
882 	unregister_blkdev(NBD_MAJOR, "nbd");
883 	kfree(nbd_dev);
884 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
885 }
886 
887 module_init(nbd_init);
888 module_exit(nbd_cleanup);
889 
890 MODULE_DESCRIPTION("Network Block Device");
891 MODULE_LICENSE("GPL");
892 
893 module_param(nbds_max, int, 0444);
894 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
895 module_param(max_part, int, 0444);
896 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
897