xref: /openbmc/linux/drivers/block/nbd.c (revision 62e7ca52)
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_SET_TIMEOUT: return "set-timeout";
82 	case NBD_SET_FLAGS: return "set-flags";
83 	case NBD_DO_IT: return "do-it";
84 	case NBD_CLEAR_SOCK: return "clear-sock";
85 	case NBD_CLEAR_QUE: return "clear-que";
86 	case NBD_PRINT_DEBUG: return "print-debug";
87 	case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
88 	case NBD_DISCONNECT: return "disconnect";
89 	case BLKROSET: return "set-read-only";
90 	case BLKFLSBUF: return "flush-buffer-cache";
91 	}
92 	return "unknown";
93 }
94 
95 static const char *nbdcmd_to_ascii(int cmd)
96 {
97 	switch (cmd) {
98 	case  NBD_CMD_READ: return "read";
99 	case NBD_CMD_WRITE: return "write";
100 	case  NBD_CMD_DISC: return "disconnect";
101 	case NBD_CMD_FLUSH: return "flush";
102 	case  NBD_CMD_TRIM: return "trim/discard";
103 	}
104 	return "invalid";
105 }
106 #endif /* NDEBUG */
107 
108 static void nbd_end_request(struct request *req)
109 {
110 	int error = req->errors ? -EIO : 0;
111 	struct request_queue *q = req->q;
112 	unsigned long flags;
113 
114 	dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
115 			req, error ? "failed" : "done");
116 
117 	spin_lock_irqsave(q->queue_lock, flags);
118 	__blk_end_request_all(req, error);
119 	spin_unlock_irqrestore(q->queue_lock, flags);
120 }
121 
122 static void sock_shutdown(struct nbd_device *nbd, int lock)
123 {
124 	/* Forcibly shutdown the socket causing all listeners
125 	 * to error
126 	 *
127 	 * FIXME: This code is duplicated from sys_shutdown, but
128 	 * there should be a more generic interface rather than
129 	 * calling socket ops directly here */
130 	if (lock)
131 		mutex_lock(&nbd->tx_lock);
132 	if (nbd->sock) {
133 		dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
134 		kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
135 		nbd->sock = NULL;
136 	}
137 	if (lock)
138 		mutex_unlock(&nbd->tx_lock);
139 }
140 
141 static void nbd_xmit_timeout(unsigned long arg)
142 {
143 	struct task_struct *task = (struct task_struct *)arg;
144 
145 	printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
146 		task->comm, task->pid);
147 	force_sig(SIGKILL, task);
148 }
149 
150 /*
151  *  Send or receive packet.
152  */
153 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
154 		int msg_flags)
155 {
156 	struct socket *sock = nbd->sock;
157 	int result;
158 	struct msghdr msg;
159 	struct kvec iov;
160 	sigset_t blocked, oldset;
161 	unsigned long pflags = current->flags;
162 
163 	if (unlikely(!sock)) {
164 		dev_err(disk_to_dev(nbd->disk),
165 			"Attempted %s on closed socket in sock_xmit\n",
166 			(send ? "send" : "recv"));
167 		return -EINVAL;
168 	}
169 
170 	/* Allow interception of SIGKILL only
171 	 * Don't allow other signals to interrupt the transmission */
172 	siginitsetinv(&blocked, sigmask(SIGKILL));
173 	sigprocmask(SIG_SETMASK, &blocked, &oldset);
174 
175 	current->flags |= PF_MEMALLOC;
176 	do {
177 		sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
178 		iov.iov_base = buf;
179 		iov.iov_len = size;
180 		msg.msg_name = NULL;
181 		msg.msg_namelen = 0;
182 		msg.msg_control = NULL;
183 		msg.msg_controllen = 0;
184 		msg.msg_flags = msg_flags | MSG_NOSIGNAL;
185 
186 		if (send) {
187 			struct timer_list ti;
188 
189 			if (nbd->xmit_timeout) {
190 				init_timer(&ti);
191 				ti.function = nbd_xmit_timeout;
192 				ti.data = (unsigned long)current;
193 				ti.expires = jiffies + nbd->xmit_timeout;
194 				add_timer(&ti);
195 			}
196 			result = kernel_sendmsg(sock, &msg, &iov, 1, size);
197 			if (nbd->xmit_timeout)
198 				del_timer_sync(&ti);
199 		} else
200 			result = kernel_recvmsg(sock, &msg, &iov, 1, size,
201 						msg.msg_flags);
202 
203 		if (signal_pending(current)) {
204 			siginfo_t info;
205 			printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
206 				task_pid_nr(current), current->comm,
207 				dequeue_signal_lock(current, &current->blocked, &info));
208 			result = -EINTR;
209 			sock_shutdown(nbd, !send);
210 			break;
211 		}
212 
213 		if (result <= 0) {
214 			if (result == 0)
215 				result = -EPIPE; /* short read */
216 			break;
217 		}
218 		size -= result;
219 		buf += result;
220 	} while (size > 0);
221 
222 	sigprocmask(SIG_SETMASK, &oldset, NULL);
223 	tsk_restore_flags(current, pflags, PF_MEMALLOC);
224 
225 	return result;
226 }
227 
228 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
229 		int flags)
230 {
231 	int result;
232 	void *kaddr = kmap(bvec->bv_page);
233 	result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
234 			   bvec->bv_len, flags);
235 	kunmap(bvec->bv_page);
236 	return result;
237 }
238 
239 /* always call with the tx_lock held */
240 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
241 {
242 	int result, flags;
243 	struct nbd_request request;
244 	unsigned long size = blk_rq_bytes(req);
245 
246 	memset(&request, 0, sizeof(request));
247 	request.magic = htonl(NBD_REQUEST_MAGIC);
248 	request.type = htonl(nbd_cmd(req));
249 
250 	if (nbd_cmd(req) != NBD_CMD_FLUSH && nbd_cmd(req) != NBD_CMD_DISC) {
251 		request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
252 		request.len = htonl(size);
253 	}
254 	memcpy(request.handle, &req, sizeof(req));
255 
256 	dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
257 			nbd->disk->disk_name, req,
258 			nbdcmd_to_ascii(nbd_cmd(req)),
259 			(unsigned long long)blk_rq_pos(req) << 9,
260 			blk_rq_bytes(req));
261 	result = sock_xmit(nbd, 1, &request, sizeof(request),
262 			(nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
263 	if (result <= 0) {
264 		dev_err(disk_to_dev(nbd->disk),
265 			"Send control failed (result %d)\n", result);
266 		goto error_out;
267 	}
268 
269 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
270 		struct req_iterator iter;
271 		struct bio_vec bvec;
272 		/*
273 		 * we are really probing at internals to determine
274 		 * whether to set MSG_MORE or not...
275 		 */
276 		rq_for_each_segment(bvec, req, iter) {
277 			flags = 0;
278 			if (!rq_iter_last(bvec, iter))
279 				flags = MSG_MORE;
280 			dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
281 					nbd->disk->disk_name, req, bvec.bv_len);
282 			result = sock_send_bvec(nbd, &bvec, flags);
283 			if (result <= 0) {
284 				dev_err(disk_to_dev(nbd->disk),
285 					"Send data failed (result %d)\n",
286 					result);
287 				goto error_out;
288 			}
289 		}
290 	}
291 	return 0;
292 
293 error_out:
294 	return -EIO;
295 }
296 
297 static struct request *nbd_find_request(struct nbd_device *nbd,
298 					struct request *xreq)
299 {
300 	struct request *req, *tmp;
301 	int err;
302 
303 	err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
304 	if (unlikely(err))
305 		goto out;
306 
307 	spin_lock(&nbd->queue_lock);
308 	list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
309 		if (req != xreq)
310 			continue;
311 		list_del_init(&req->queuelist);
312 		spin_unlock(&nbd->queue_lock);
313 		return req;
314 	}
315 	spin_unlock(&nbd->queue_lock);
316 
317 	err = -ENOENT;
318 
319 out:
320 	return ERR_PTR(err);
321 }
322 
323 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
324 {
325 	int result;
326 	void *kaddr = kmap(bvec->bv_page);
327 	result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
328 			MSG_WAITALL);
329 	kunmap(bvec->bv_page);
330 	return result;
331 }
332 
333 /* NULL returned = something went wrong, inform userspace */
334 static struct request *nbd_read_stat(struct nbd_device *nbd)
335 {
336 	int result;
337 	struct nbd_reply reply;
338 	struct request *req;
339 
340 	reply.magic = 0;
341 	result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
342 	if (result <= 0) {
343 		dev_err(disk_to_dev(nbd->disk),
344 			"Receive control failed (result %d)\n", result);
345 		goto harderror;
346 	}
347 
348 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
349 		dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
350 				(unsigned long)ntohl(reply.magic));
351 		result = -EPROTO;
352 		goto harderror;
353 	}
354 
355 	req = nbd_find_request(nbd, *(struct request **)reply.handle);
356 	if (IS_ERR(req)) {
357 		result = PTR_ERR(req);
358 		if (result != -ENOENT)
359 			goto harderror;
360 
361 		dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
362 			reply.handle);
363 		result = -EBADR;
364 		goto harderror;
365 	}
366 
367 	if (ntohl(reply.error)) {
368 		dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
369 			ntohl(reply.error));
370 		req->errors++;
371 		return req;
372 	}
373 
374 	dprintk(DBG_RX, "%s: request %p: got reply\n",
375 			nbd->disk->disk_name, req);
376 	if (nbd_cmd(req) == NBD_CMD_READ) {
377 		struct req_iterator iter;
378 		struct bio_vec bvec;
379 
380 		rq_for_each_segment(bvec, req, iter) {
381 			result = sock_recv_bvec(nbd, &bvec);
382 			if (result <= 0) {
383 				dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
384 					result);
385 				req->errors++;
386 				return req;
387 			}
388 			dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
389 				nbd->disk->disk_name, req, bvec.bv_len);
390 		}
391 	}
392 	return req;
393 harderror:
394 	nbd->harderror = result;
395 	return NULL;
396 }
397 
398 static ssize_t pid_show(struct device *dev,
399 			struct device_attribute *attr, char *buf)
400 {
401 	struct gendisk *disk = dev_to_disk(dev);
402 
403 	return sprintf(buf, "%ld\n",
404 		(long) ((struct nbd_device *)disk->private_data)->pid);
405 }
406 
407 static struct device_attribute pid_attr = {
408 	.attr = { .name = "pid", .mode = S_IRUGO},
409 	.show = pid_show,
410 };
411 
412 static int nbd_do_it(struct nbd_device *nbd)
413 {
414 	struct request *req;
415 	int ret;
416 
417 	BUG_ON(nbd->magic != NBD_MAGIC);
418 
419 	sk_set_memalloc(nbd->sock->sk);
420 	nbd->pid = task_pid_nr(current);
421 	ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
422 	if (ret) {
423 		dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
424 		nbd->pid = 0;
425 		return ret;
426 	}
427 
428 	while ((req = nbd_read_stat(nbd)) != NULL)
429 		nbd_end_request(req);
430 
431 	device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
432 	nbd->pid = 0;
433 	return 0;
434 }
435 
436 static void nbd_clear_que(struct nbd_device *nbd)
437 {
438 	struct request *req;
439 
440 	BUG_ON(nbd->magic != NBD_MAGIC);
441 
442 	/*
443 	 * Because we have set nbd->sock to NULL under the tx_lock, all
444 	 * modifications to the list must have completed by now.  For
445 	 * the same reason, the active_req must be NULL.
446 	 *
447 	 * As a consequence, we don't need to take the spin lock while
448 	 * purging the list here.
449 	 */
450 	BUG_ON(nbd->sock);
451 	BUG_ON(nbd->active_req);
452 
453 	while (!list_empty(&nbd->queue_head)) {
454 		req = list_entry(nbd->queue_head.next, struct request,
455 				 queuelist);
456 		list_del_init(&req->queuelist);
457 		req->errors++;
458 		nbd_end_request(req);
459 	}
460 
461 	while (!list_empty(&nbd->waiting_queue)) {
462 		req = list_entry(nbd->waiting_queue.next, struct request,
463 				 queuelist);
464 		list_del_init(&req->queuelist);
465 		req->errors++;
466 		nbd_end_request(req);
467 	}
468 }
469 
470 
471 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
472 {
473 	if (req->cmd_type != REQ_TYPE_FS)
474 		goto error_out;
475 
476 	nbd_cmd(req) = NBD_CMD_READ;
477 	if (rq_data_dir(req) == WRITE) {
478 		if ((req->cmd_flags & REQ_DISCARD)) {
479 			WARN_ON(!(nbd->flags & NBD_FLAG_SEND_TRIM));
480 			nbd_cmd(req) = NBD_CMD_TRIM;
481 		} else
482 			nbd_cmd(req) = NBD_CMD_WRITE;
483 		if (nbd->flags & NBD_FLAG_READ_ONLY) {
484 			dev_err(disk_to_dev(nbd->disk),
485 				"Write on read-only\n");
486 			goto error_out;
487 		}
488 	}
489 
490 	if (req->cmd_flags & REQ_FLUSH) {
491 		BUG_ON(unlikely(blk_rq_sectors(req)));
492 		nbd_cmd(req) = NBD_CMD_FLUSH;
493 	}
494 
495 	req->errors = 0;
496 
497 	mutex_lock(&nbd->tx_lock);
498 	if (unlikely(!nbd->sock)) {
499 		mutex_unlock(&nbd->tx_lock);
500 		dev_err(disk_to_dev(nbd->disk),
501 			"Attempted send on closed socket\n");
502 		goto error_out;
503 	}
504 
505 	nbd->active_req = req;
506 
507 	if (nbd_send_req(nbd, req) != 0) {
508 		dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
509 		req->errors++;
510 		nbd_end_request(req);
511 	} else {
512 		spin_lock(&nbd->queue_lock);
513 		list_add_tail(&req->queuelist, &nbd->queue_head);
514 		spin_unlock(&nbd->queue_lock);
515 	}
516 
517 	nbd->active_req = NULL;
518 	mutex_unlock(&nbd->tx_lock);
519 	wake_up_all(&nbd->active_wq);
520 
521 	return;
522 
523 error_out:
524 	req->errors++;
525 	nbd_end_request(req);
526 }
527 
528 static int nbd_thread(void *data)
529 {
530 	struct nbd_device *nbd = data;
531 	struct request *req;
532 
533 	set_user_nice(current, MIN_NICE);
534 	while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
535 		/* wait for something to do */
536 		wait_event_interruptible(nbd->waiting_wq,
537 					 kthread_should_stop() ||
538 					 !list_empty(&nbd->waiting_queue));
539 
540 		/* extract request */
541 		if (list_empty(&nbd->waiting_queue))
542 			continue;
543 
544 		spin_lock_irq(&nbd->queue_lock);
545 		req = list_entry(nbd->waiting_queue.next, struct request,
546 				 queuelist);
547 		list_del_init(&req->queuelist);
548 		spin_unlock_irq(&nbd->queue_lock);
549 
550 		/* handle request */
551 		nbd_handle_req(nbd, req);
552 	}
553 	return 0;
554 }
555 
556 /*
557  * We always wait for result of write, for now. It would be nice to make it optional
558  * in future
559  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
560  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
561  */
562 
563 static void do_nbd_request(struct request_queue *q)
564 		__releases(q->queue_lock) __acquires(q->queue_lock)
565 {
566 	struct request *req;
567 
568 	while ((req = blk_fetch_request(q)) != NULL) {
569 		struct nbd_device *nbd;
570 
571 		spin_unlock_irq(q->queue_lock);
572 
573 		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
574 				req->rq_disk->disk_name, req, req->cmd_type);
575 
576 		nbd = req->rq_disk->private_data;
577 
578 		BUG_ON(nbd->magic != NBD_MAGIC);
579 
580 		if (unlikely(!nbd->sock)) {
581 			dev_err(disk_to_dev(nbd->disk),
582 				"Attempted send on closed socket\n");
583 			req->errors++;
584 			nbd_end_request(req);
585 			spin_lock_irq(q->queue_lock);
586 			continue;
587 		}
588 
589 		spin_lock_irq(&nbd->queue_lock);
590 		list_add_tail(&req->queuelist, &nbd->waiting_queue);
591 		spin_unlock_irq(&nbd->queue_lock);
592 
593 		wake_up(&nbd->waiting_wq);
594 
595 		spin_lock_irq(q->queue_lock);
596 	}
597 }
598 
599 /* Must be called with tx_lock held */
600 
601 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
602 		       unsigned int cmd, unsigned long arg)
603 {
604 	switch (cmd) {
605 	case NBD_DISCONNECT: {
606 		struct request sreq;
607 
608 		dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
609 		if (!nbd->sock)
610 			return -EINVAL;
611 
612 		mutex_unlock(&nbd->tx_lock);
613 		fsync_bdev(bdev);
614 		mutex_lock(&nbd->tx_lock);
615 		blk_rq_init(NULL, &sreq);
616 		sreq.cmd_type = REQ_TYPE_SPECIAL;
617 		nbd_cmd(&sreq) = NBD_CMD_DISC;
618 
619 		/* Check again after getting mutex back.  */
620 		if (!nbd->sock)
621 			return -EINVAL;
622 
623 		nbd->disconnect = 1;
624 
625 		nbd_send_req(nbd, &sreq);
626 		return 0;
627 	}
628 
629 	case NBD_CLEAR_SOCK: {
630 		struct socket *sock = nbd->sock;
631 		nbd->sock = NULL;
632 		nbd_clear_que(nbd);
633 		BUG_ON(!list_empty(&nbd->queue_head));
634 		BUG_ON(!list_empty(&nbd->waiting_queue));
635 		kill_bdev(bdev);
636 		if (sock)
637 			sockfd_put(sock);
638 		return 0;
639 	}
640 
641 	case NBD_SET_SOCK: {
642 		struct socket *sock;
643 		int err;
644 		if (nbd->sock)
645 			return -EBUSY;
646 		sock = sockfd_lookup(arg, &err);
647 		if (sock) {
648 			nbd->sock = sock;
649 			if (max_part > 0)
650 				bdev->bd_invalidated = 1;
651 			nbd->disconnect = 0; /* we're connected now */
652 			return 0;
653 		}
654 		return -EINVAL;
655 	}
656 
657 	case NBD_SET_BLKSIZE:
658 		nbd->blksize = arg;
659 		nbd->bytesize &= ~(nbd->blksize-1);
660 		bdev->bd_inode->i_size = nbd->bytesize;
661 		set_blocksize(bdev, nbd->blksize);
662 		set_capacity(nbd->disk, nbd->bytesize >> 9);
663 		return 0;
664 
665 	case NBD_SET_SIZE:
666 		nbd->bytesize = arg & ~(nbd->blksize-1);
667 		bdev->bd_inode->i_size = nbd->bytesize;
668 		set_blocksize(bdev, nbd->blksize);
669 		set_capacity(nbd->disk, nbd->bytesize >> 9);
670 		return 0;
671 
672 	case NBD_SET_TIMEOUT:
673 		nbd->xmit_timeout = arg * HZ;
674 		return 0;
675 
676 	case NBD_SET_FLAGS:
677 		nbd->flags = arg;
678 		return 0;
679 
680 	case NBD_SET_SIZE_BLOCKS:
681 		nbd->bytesize = ((u64) arg) * nbd->blksize;
682 		bdev->bd_inode->i_size = nbd->bytesize;
683 		set_blocksize(bdev, nbd->blksize);
684 		set_capacity(nbd->disk, nbd->bytesize >> 9);
685 		return 0;
686 
687 	case NBD_DO_IT: {
688 		struct task_struct *thread;
689 		struct socket *sock;
690 		int error;
691 
692 		if (nbd->pid)
693 			return -EBUSY;
694 		if (!nbd->sock)
695 			return -EINVAL;
696 
697 		mutex_unlock(&nbd->tx_lock);
698 
699 		if (nbd->flags & NBD_FLAG_READ_ONLY)
700 			set_device_ro(bdev, true);
701 		if (nbd->flags & NBD_FLAG_SEND_TRIM)
702 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
703 				nbd->disk->queue);
704 		if (nbd->flags & NBD_FLAG_SEND_FLUSH)
705 			blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
706 		else
707 			blk_queue_flush(nbd->disk->queue, 0);
708 
709 		thread = kthread_create(nbd_thread, nbd, "%s",
710 					nbd->disk->disk_name);
711 		if (IS_ERR(thread)) {
712 			mutex_lock(&nbd->tx_lock);
713 			return PTR_ERR(thread);
714 		}
715 		wake_up_process(thread);
716 		error = nbd_do_it(nbd);
717 		kthread_stop(thread);
718 
719 		mutex_lock(&nbd->tx_lock);
720 		if (error)
721 			return error;
722 		sock_shutdown(nbd, 0);
723 		sock = nbd->sock;
724 		nbd->sock = NULL;
725 		nbd_clear_que(nbd);
726 		dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
727 		kill_bdev(bdev);
728 		queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
729 		set_device_ro(bdev, false);
730 		if (sock)
731 			sockfd_put(sock);
732 		nbd->flags = 0;
733 		nbd->bytesize = 0;
734 		bdev->bd_inode->i_size = 0;
735 		set_capacity(nbd->disk, 0);
736 		if (max_part > 0)
737 			ioctl_by_bdev(bdev, BLKRRPART, 0);
738 		if (nbd->disconnect) /* user requested, ignore socket errors */
739 			return 0;
740 		return nbd->harderror;
741 	}
742 
743 	case NBD_CLEAR_QUE:
744 		/*
745 		 * This is for compatibility only.  The queue is always cleared
746 		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
747 		 */
748 		return 0;
749 
750 	case NBD_PRINT_DEBUG:
751 		dev_info(disk_to_dev(nbd->disk),
752 			"next = %p, prev = %p, head = %p\n",
753 			nbd->queue_head.next, nbd->queue_head.prev,
754 			&nbd->queue_head);
755 		return 0;
756 	}
757 	return -ENOTTY;
758 }
759 
760 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
761 		     unsigned int cmd, unsigned long arg)
762 {
763 	struct nbd_device *nbd = bdev->bd_disk->private_data;
764 	int error;
765 
766 	if (!capable(CAP_SYS_ADMIN))
767 		return -EPERM;
768 
769 	BUG_ON(nbd->magic != NBD_MAGIC);
770 
771 	/* Anyone capable of this syscall can do *real bad* things */
772 	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
773 		nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
774 
775 	mutex_lock(&nbd->tx_lock);
776 	error = __nbd_ioctl(bdev, nbd, cmd, arg);
777 	mutex_unlock(&nbd->tx_lock);
778 
779 	return error;
780 }
781 
782 static const struct block_device_operations nbd_fops =
783 {
784 	.owner =	THIS_MODULE,
785 	.ioctl =	nbd_ioctl,
786 };
787 
788 /*
789  * And here should be modules and kernel interface
790  *  (Just smiley confuses emacs :-)
791  */
792 
793 static int __init nbd_init(void)
794 {
795 	int err = -ENOMEM;
796 	int i;
797 	int part_shift;
798 
799 	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
800 
801 	if (max_part < 0) {
802 		printk(KERN_ERR "nbd: max_part must be >= 0\n");
803 		return -EINVAL;
804 	}
805 
806 	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
807 	if (!nbd_dev)
808 		return -ENOMEM;
809 
810 	part_shift = 0;
811 	if (max_part > 0) {
812 		part_shift = fls(max_part);
813 
814 		/*
815 		 * Adjust max_part according to part_shift as it is exported
816 		 * to user space so that user can know the max number of
817 		 * partition kernel should be able to manage.
818 		 *
819 		 * Note that -1 is required because partition 0 is reserved
820 		 * for the whole disk.
821 		 */
822 		max_part = (1UL << part_shift) - 1;
823 	}
824 
825 	if ((1UL << part_shift) > DISK_MAX_PARTS)
826 		return -EINVAL;
827 
828 	if (nbds_max > 1UL << (MINORBITS - part_shift))
829 		return -EINVAL;
830 
831 	for (i = 0; i < nbds_max; i++) {
832 		struct gendisk *disk = alloc_disk(1 << part_shift);
833 		if (!disk)
834 			goto out;
835 		nbd_dev[i].disk = disk;
836 		/*
837 		 * The new linux 2.5 block layer implementation requires
838 		 * every gendisk to have its very own request_queue struct.
839 		 * These structs are big so we dynamically allocate them.
840 		 */
841 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
842 		if (!disk->queue) {
843 			put_disk(disk);
844 			goto out;
845 		}
846 		/*
847 		 * Tell the block layer that we are not a rotational device
848 		 */
849 		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
850 		disk->queue->limits.discard_granularity = 512;
851 		disk->queue->limits.max_discard_sectors = UINT_MAX;
852 		disk->queue->limits.discard_zeroes_data = 0;
853 		blk_queue_max_hw_sectors(disk->queue, 65536);
854 		disk->queue->limits.max_sectors = 256;
855 	}
856 
857 	if (register_blkdev(NBD_MAJOR, "nbd")) {
858 		err = -EIO;
859 		goto out;
860 	}
861 
862 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
863 	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
864 
865 	for (i = 0; i < nbds_max; i++) {
866 		struct gendisk *disk = nbd_dev[i].disk;
867 		nbd_dev[i].magic = NBD_MAGIC;
868 		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
869 		spin_lock_init(&nbd_dev[i].queue_lock);
870 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
871 		mutex_init(&nbd_dev[i].tx_lock);
872 		init_waitqueue_head(&nbd_dev[i].active_wq);
873 		init_waitqueue_head(&nbd_dev[i].waiting_wq);
874 		nbd_dev[i].blksize = 1024;
875 		nbd_dev[i].bytesize = 0;
876 		disk->major = NBD_MAJOR;
877 		disk->first_minor = i << part_shift;
878 		disk->fops = &nbd_fops;
879 		disk->private_data = &nbd_dev[i];
880 		sprintf(disk->disk_name, "nbd%d", i);
881 		set_capacity(disk, 0);
882 		add_disk(disk);
883 	}
884 
885 	return 0;
886 out:
887 	while (i--) {
888 		blk_cleanup_queue(nbd_dev[i].disk->queue);
889 		put_disk(nbd_dev[i].disk);
890 	}
891 	kfree(nbd_dev);
892 	return err;
893 }
894 
895 static void __exit nbd_cleanup(void)
896 {
897 	int i;
898 	for (i = 0; i < nbds_max; i++) {
899 		struct gendisk *disk = nbd_dev[i].disk;
900 		nbd_dev[i].magic = 0;
901 		if (disk) {
902 			del_gendisk(disk);
903 			blk_cleanup_queue(disk->queue);
904 			put_disk(disk);
905 		}
906 	}
907 	unregister_blkdev(NBD_MAJOR, "nbd");
908 	kfree(nbd_dev);
909 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
910 }
911 
912 module_init(nbd_init);
913 module_exit(nbd_cleanup);
914 
915 MODULE_DESCRIPTION("Network Block Device");
916 MODULE_LICENSE("GPL");
917 
918 module_param(nbds_max, int, 0444);
919 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
920 module_param(max_part, int, 0444);
921 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
922 #ifndef NDEBUG
923 module_param(debugflags, int, 0644);
924 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
925 #endif
926