xref: /openbmc/linux/drivers/block/nbd.c (revision 12eb4683)
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 	request.magic = htonl(NBD_REQUEST_MAGIC);
247 	request.type = htonl(nbd_cmd(req));
248 
249 	if (nbd_cmd(req) == NBD_CMD_FLUSH) {
250 		/* Other values are reserved for FLUSH requests.  */
251 		request.from = 0;
252 		request.len = 0;
253 	} else {
254 		request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
255 		request.len = htonl(size);
256 	}
257 	memcpy(request.handle, &req, sizeof(req));
258 
259 	dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
260 			nbd->disk->disk_name, req,
261 			nbdcmd_to_ascii(nbd_cmd(req)),
262 			(unsigned long long)blk_rq_pos(req) << 9,
263 			blk_rq_bytes(req));
264 	result = sock_xmit(nbd, 1, &request, sizeof(request),
265 			(nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
266 	if (result <= 0) {
267 		dev_err(disk_to_dev(nbd->disk),
268 			"Send control failed (result %d)\n", result);
269 		goto error_out;
270 	}
271 
272 	if (nbd_cmd(req) == NBD_CMD_WRITE) {
273 		struct req_iterator iter;
274 		struct bio_vec *bvec;
275 		/*
276 		 * we are really probing at internals to determine
277 		 * whether to set MSG_MORE or not...
278 		 */
279 		rq_for_each_segment(bvec, req, iter) {
280 			flags = 0;
281 			if (!rq_iter_last(req, iter))
282 				flags = MSG_MORE;
283 			dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
284 					nbd->disk->disk_name, req, bvec->bv_len);
285 			result = sock_send_bvec(nbd, bvec, flags);
286 			if (result <= 0) {
287 				dev_err(disk_to_dev(nbd->disk),
288 					"Send data failed (result %d)\n",
289 					result);
290 				goto error_out;
291 			}
292 		}
293 	}
294 	return 0;
295 
296 error_out:
297 	return -EIO;
298 }
299 
300 static struct request *nbd_find_request(struct nbd_device *nbd,
301 					struct request *xreq)
302 {
303 	struct request *req, *tmp;
304 	int err;
305 
306 	err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
307 	if (unlikely(err))
308 		goto out;
309 
310 	spin_lock(&nbd->queue_lock);
311 	list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
312 		if (req != xreq)
313 			continue;
314 		list_del_init(&req->queuelist);
315 		spin_unlock(&nbd->queue_lock);
316 		return req;
317 	}
318 	spin_unlock(&nbd->queue_lock);
319 
320 	err = -ENOENT;
321 
322 out:
323 	return ERR_PTR(err);
324 }
325 
326 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
327 {
328 	int result;
329 	void *kaddr = kmap(bvec->bv_page);
330 	result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
331 			MSG_WAITALL);
332 	kunmap(bvec->bv_page);
333 	return result;
334 }
335 
336 /* NULL returned = something went wrong, inform userspace */
337 static struct request *nbd_read_stat(struct nbd_device *nbd)
338 {
339 	int result;
340 	struct nbd_reply reply;
341 	struct request *req;
342 
343 	reply.magic = 0;
344 	result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
345 	if (result <= 0) {
346 		dev_err(disk_to_dev(nbd->disk),
347 			"Receive control failed (result %d)\n", result);
348 		goto harderror;
349 	}
350 
351 	if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
352 		dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
353 				(unsigned long)ntohl(reply.magic));
354 		result = -EPROTO;
355 		goto harderror;
356 	}
357 
358 	req = nbd_find_request(nbd, *(struct request **)reply.handle);
359 	if (IS_ERR(req)) {
360 		result = PTR_ERR(req);
361 		if (result != -ENOENT)
362 			goto harderror;
363 
364 		dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
365 			reply.handle);
366 		result = -EBADR;
367 		goto harderror;
368 	}
369 
370 	if (ntohl(reply.error)) {
371 		dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
372 			ntohl(reply.error));
373 		req->errors++;
374 		return req;
375 	}
376 
377 	dprintk(DBG_RX, "%s: request %p: got reply\n",
378 			nbd->disk->disk_name, req);
379 	if (nbd_cmd(req) == NBD_CMD_READ) {
380 		struct req_iterator iter;
381 		struct bio_vec *bvec;
382 
383 		rq_for_each_segment(bvec, req, iter) {
384 			result = sock_recv_bvec(nbd, bvec);
385 			if (result <= 0) {
386 				dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
387 					result);
388 				req->errors++;
389 				return req;
390 			}
391 			dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
392 				nbd->disk->disk_name, req, bvec->bv_len);
393 		}
394 	}
395 	return req;
396 harderror:
397 	nbd->harderror = result;
398 	return NULL;
399 }
400 
401 static ssize_t pid_show(struct device *dev,
402 			struct device_attribute *attr, char *buf)
403 {
404 	struct gendisk *disk = dev_to_disk(dev);
405 
406 	return sprintf(buf, "%ld\n",
407 		(long) ((struct nbd_device *)disk->private_data)->pid);
408 }
409 
410 static struct device_attribute pid_attr = {
411 	.attr = { .name = "pid", .mode = S_IRUGO},
412 	.show = pid_show,
413 };
414 
415 static int nbd_do_it(struct nbd_device *nbd)
416 {
417 	struct request *req;
418 	int ret;
419 
420 	BUG_ON(nbd->magic != NBD_MAGIC);
421 
422 	sk_set_memalloc(nbd->sock->sk);
423 	nbd->pid = task_pid_nr(current);
424 	ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
425 	if (ret) {
426 		dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
427 		nbd->pid = 0;
428 		return ret;
429 	}
430 
431 	while ((req = nbd_read_stat(nbd)) != NULL)
432 		nbd_end_request(req);
433 
434 	device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
435 	nbd->pid = 0;
436 	return 0;
437 }
438 
439 static void nbd_clear_que(struct nbd_device *nbd)
440 {
441 	struct request *req;
442 
443 	BUG_ON(nbd->magic != NBD_MAGIC);
444 
445 	/*
446 	 * Because we have set nbd->sock to NULL under the tx_lock, all
447 	 * modifications to the list must have completed by now.  For
448 	 * the same reason, the active_req must be NULL.
449 	 *
450 	 * As a consequence, we don't need to take the spin lock while
451 	 * purging the list here.
452 	 */
453 	BUG_ON(nbd->sock);
454 	BUG_ON(nbd->active_req);
455 
456 	while (!list_empty(&nbd->queue_head)) {
457 		req = list_entry(nbd->queue_head.next, struct request,
458 				 queuelist);
459 		list_del_init(&req->queuelist);
460 		req->errors++;
461 		nbd_end_request(req);
462 	}
463 
464 	while (!list_empty(&nbd->waiting_queue)) {
465 		req = list_entry(nbd->waiting_queue.next, struct request,
466 				 queuelist);
467 		list_del_init(&req->queuelist);
468 		req->errors++;
469 		nbd_end_request(req);
470 	}
471 }
472 
473 
474 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
475 {
476 	if (req->cmd_type != REQ_TYPE_FS)
477 		goto error_out;
478 
479 	nbd_cmd(req) = NBD_CMD_READ;
480 	if (rq_data_dir(req) == WRITE) {
481 		if ((req->cmd_flags & REQ_DISCARD)) {
482 			WARN_ON(!(nbd->flags & NBD_FLAG_SEND_TRIM));
483 			nbd_cmd(req) = NBD_CMD_TRIM;
484 		} else
485 			nbd_cmd(req) = NBD_CMD_WRITE;
486 		if (nbd->flags & NBD_FLAG_READ_ONLY) {
487 			dev_err(disk_to_dev(nbd->disk),
488 				"Write on read-only\n");
489 			goto error_out;
490 		}
491 	}
492 
493 	if (req->cmd_flags & REQ_FLUSH) {
494 		BUG_ON(unlikely(blk_rq_sectors(req)));
495 		nbd_cmd(req) = NBD_CMD_FLUSH;
496 	}
497 
498 	req->errors = 0;
499 
500 	mutex_lock(&nbd->tx_lock);
501 	if (unlikely(!nbd->sock)) {
502 		mutex_unlock(&nbd->tx_lock);
503 		dev_err(disk_to_dev(nbd->disk),
504 			"Attempted send on closed socket\n");
505 		goto error_out;
506 	}
507 
508 	nbd->active_req = req;
509 
510 	if (nbd_send_req(nbd, req) != 0) {
511 		dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
512 		req->errors++;
513 		nbd_end_request(req);
514 	} else {
515 		spin_lock(&nbd->queue_lock);
516 		list_add_tail(&req->queuelist, &nbd->queue_head);
517 		spin_unlock(&nbd->queue_lock);
518 	}
519 
520 	nbd->active_req = NULL;
521 	mutex_unlock(&nbd->tx_lock);
522 	wake_up_all(&nbd->active_wq);
523 
524 	return;
525 
526 error_out:
527 	req->errors++;
528 	nbd_end_request(req);
529 }
530 
531 static int nbd_thread(void *data)
532 {
533 	struct nbd_device *nbd = data;
534 	struct request *req;
535 
536 	set_user_nice(current, -20);
537 	while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
538 		/* wait for something to do */
539 		wait_event_interruptible(nbd->waiting_wq,
540 					 kthread_should_stop() ||
541 					 !list_empty(&nbd->waiting_queue));
542 
543 		/* extract request */
544 		if (list_empty(&nbd->waiting_queue))
545 			continue;
546 
547 		spin_lock_irq(&nbd->queue_lock);
548 		req = list_entry(nbd->waiting_queue.next, struct request,
549 				 queuelist);
550 		list_del_init(&req->queuelist);
551 		spin_unlock_irq(&nbd->queue_lock);
552 
553 		/* handle request */
554 		nbd_handle_req(nbd, req);
555 	}
556 	return 0;
557 }
558 
559 /*
560  * We always wait for result of write, for now. It would be nice to make it optional
561  * in future
562  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
563  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
564  */
565 
566 static void do_nbd_request(struct request_queue *q)
567 		__releases(q->queue_lock) __acquires(q->queue_lock)
568 {
569 	struct request *req;
570 
571 	while ((req = blk_fetch_request(q)) != NULL) {
572 		struct nbd_device *nbd;
573 
574 		spin_unlock_irq(q->queue_lock);
575 
576 		dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
577 				req->rq_disk->disk_name, req, req->cmd_type);
578 
579 		nbd = req->rq_disk->private_data;
580 
581 		BUG_ON(nbd->magic != NBD_MAGIC);
582 
583 		if (unlikely(!nbd->sock)) {
584 			dev_err(disk_to_dev(nbd->disk),
585 				"Attempted send on closed socket\n");
586 			req->errors++;
587 			nbd_end_request(req);
588 			spin_lock_irq(q->queue_lock);
589 			continue;
590 		}
591 
592 		spin_lock_irq(&nbd->queue_lock);
593 		list_add_tail(&req->queuelist, &nbd->waiting_queue);
594 		spin_unlock_irq(&nbd->queue_lock);
595 
596 		wake_up(&nbd->waiting_wq);
597 
598 		spin_lock_irq(q->queue_lock);
599 	}
600 }
601 
602 /* Must be called with tx_lock held */
603 
604 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
605 		       unsigned int cmd, unsigned long arg)
606 {
607 	switch (cmd) {
608 	case NBD_DISCONNECT: {
609 		struct request sreq;
610 
611 		dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
612 		if (!nbd->sock)
613 			return -EINVAL;
614 
615 		mutex_unlock(&nbd->tx_lock);
616 		fsync_bdev(bdev);
617 		mutex_lock(&nbd->tx_lock);
618 		blk_rq_init(NULL, &sreq);
619 		sreq.cmd_type = REQ_TYPE_SPECIAL;
620 		nbd_cmd(&sreq) = NBD_CMD_DISC;
621 
622 		/* Check again after getting mutex back.  */
623 		if (!nbd->sock)
624 			return -EINVAL;
625 
626 		nbd->disconnect = 1;
627 
628 		nbd_send_req(nbd, &sreq);
629 		return 0;
630 	}
631 
632 	case NBD_CLEAR_SOCK: {
633 		struct file *file;
634 
635 		nbd->sock = NULL;
636 		file = nbd->file;
637 		nbd->file = NULL;
638 		nbd_clear_que(nbd);
639 		BUG_ON(!list_empty(&nbd->queue_head));
640 		BUG_ON(!list_empty(&nbd->waiting_queue));
641 		kill_bdev(bdev);
642 		if (file)
643 			fput(file);
644 		return 0;
645 	}
646 
647 	case NBD_SET_SOCK: {
648 		struct file *file;
649 		if (nbd->file)
650 			return -EBUSY;
651 		file = fget(arg);
652 		if (file) {
653 			struct inode *inode = file_inode(file);
654 			if (S_ISSOCK(inode->i_mode)) {
655 				nbd->file = file;
656 				nbd->sock = SOCKET_I(inode);
657 				if (max_part > 0)
658 					bdev->bd_invalidated = 1;
659 				nbd->disconnect = 0; /* we're connected now */
660 				return 0;
661 			} else {
662 				fput(file);
663 			}
664 		}
665 		return -EINVAL;
666 	}
667 
668 	case NBD_SET_BLKSIZE:
669 		nbd->blksize = arg;
670 		nbd->bytesize &= ~(nbd->blksize-1);
671 		bdev->bd_inode->i_size = nbd->bytesize;
672 		set_blocksize(bdev, nbd->blksize);
673 		set_capacity(nbd->disk, nbd->bytesize >> 9);
674 		return 0;
675 
676 	case NBD_SET_SIZE:
677 		nbd->bytesize = arg & ~(nbd->blksize-1);
678 		bdev->bd_inode->i_size = nbd->bytesize;
679 		set_blocksize(bdev, nbd->blksize);
680 		set_capacity(nbd->disk, nbd->bytesize >> 9);
681 		return 0;
682 
683 	case NBD_SET_TIMEOUT:
684 		nbd->xmit_timeout = arg * HZ;
685 		return 0;
686 
687 	case NBD_SET_FLAGS:
688 		nbd->flags = arg;
689 		return 0;
690 
691 	case NBD_SET_SIZE_BLOCKS:
692 		nbd->bytesize = ((u64) arg) * nbd->blksize;
693 		bdev->bd_inode->i_size = nbd->bytesize;
694 		set_blocksize(bdev, nbd->blksize);
695 		set_capacity(nbd->disk, nbd->bytesize >> 9);
696 		return 0;
697 
698 	case NBD_DO_IT: {
699 		struct task_struct *thread;
700 		struct file *file;
701 		int error;
702 
703 		if (nbd->pid)
704 			return -EBUSY;
705 		if (!nbd->file)
706 			return -EINVAL;
707 
708 		mutex_unlock(&nbd->tx_lock);
709 
710 		if (nbd->flags & NBD_FLAG_READ_ONLY)
711 			set_device_ro(bdev, true);
712 		if (nbd->flags & NBD_FLAG_SEND_TRIM)
713 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
714 				nbd->disk->queue);
715 		if (nbd->flags & NBD_FLAG_SEND_FLUSH)
716 			blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
717 		else
718 			blk_queue_flush(nbd->disk->queue, 0);
719 
720 		thread = kthread_create(nbd_thread, nbd, "%s",
721 					nbd->disk->disk_name);
722 		if (IS_ERR(thread)) {
723 			mutex_lock(&nbd->tx_lock);
724 			return PTR_ERR(thread);
725 		}
726 		wake_up_process(thread);
727 		error = nbd_do_it(nbd);
728 		kthread_stop(thread);
729 
730 		mutex_lock(&nbd->tx_lock);
731 		if (error)
732 			return error;
733 		sock_shutdown(nbd, 0);
734 		file = nbd->file;
735 		nbd->file = NULL;
736 		nbd_clear_que(nbd);
737 		dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
738 		kill_bdev(bdev);
739 		queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
740 		set_device_ro(bdev, false);
741 		if (file)
742 			fput(file);
743 		nbd->flags = 0;
744 		nbd->bytesize = 0;
745 		bdev->bd_inode->i_size = 0;
746 		set_capacity(nbd->disk, 0);
747 		if (max_part > 0)
748 			ioctl_by_bdev(bdev, BLKRRPART, 0);
749 		if (nbd->disconnect) /* user requested, ignore socket errors */
750 			return 0;
751 		return nbd->harderror;
752 	}
753 
754 	case NBD_CLEAR_QUE:
755 		/*
756 		 * This is for compatibility only.  The queue is always cleared
757 		 * by NBD_DO_IT or NBD_CLEAR_SOCK.
758 		 */
759 		return 0;
760 
761 	case NBD_PRINT_DEBUG:
762 		dev_info(disk_to_dev(nbd->disk),
763 			"next = %p, prev = %p, head = %p\n",
764 			nbd->queue_head.next, nbd->queue_head.prev,
765 			&nbd->queue_head);
766 		return 0;
767 	}
768 	return -ENOTTY;
769 }
770 
771 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
772 		     unsigned int cmd, unsigned long arg)
773 {
774 	struct nbd_device *nbd = bdev->bd_disk->private_data;
775 	int error;
776 
777 	if (!capable(CAP_SYS_ADMIN))
778 		return -EPERM;
779 
780 	BUG_ON(nbd->magic != NBD_MAGIC);
781 
782 	/* Anyone capable of this syscall can do *real bad* things */
783 	dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
784 		nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
785 
786 	mutex_lock(&nbd->tx_lock);
787 	error = __nbd_ioctl(bdev, nbd, cmd, arg);
788 	mutex_unlock(&nbd->tx_lock);
789 
790 	return error;
791 }
792 
793 static const struct block_device_operations nbd_fops =
794 {
795 	.owner =	THIS_MODULE,
796 	.ioctl =	nbd_ioctl,
797 };
798 
799 /*
800  * And here should be modules and kernel interface
801  *  (Just smiley confuses emacs :-)
802  */
803 
804 static int __init nbd_init(void)
805 {
806 	int err = -ENOMEM;
807 	int i;
808 	int part_shift;
809 
810 	BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
811 
812 	if (max_part < 0) {
813 		printk(KERN_ERR "nbd: max_part must be >= 0\n");
814 		return -EINVAL;
815 	}
816 
817 	nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
818 	if (!nbd_dev)
819 		return -ENOMEM;
820 
821 	part_shift = 0;
822 	if (max_part > 0) {
823 		part_shift = fls(max_part);
824 
825 		/*
826 		 * Adjust max_part according to part_shift as it is exported
827 		 * to user space so that user can know the max number of
828 		 * partition kernel should be able to manage.
829 		 *
830 		 * Note that -1 is required because partition 0 is reserved
831 		 * for the whole disk.
832 		 */
833 		max_part = (1UL << part_shift) - 1;
834 	}
835 
836 	if ((1UL << part_shift) > DISK_MAX_PARTS)
837 		return -EINVAL;
838 
839 	if (nbds_max > 1UL << (MINORBITS - part_shift))
840 		return -EINVAL;
841 
842 	for (i = 0; i < nbds_max; i++) {
843 		struct gendisk *disk = alloc_disk(1 << part_shift);
844 		if (!disk)
845 			goto out;
846 		nbd_dev[i].disk = disk;
847 		/*
848 		 * The new linux 2.5 block layer implementation requires
849 		 * every gendisk to have its very own request_queue struct.
850 		 * These structs are big so we dynamically allocate them.
851 		 */
852 		disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
853 		if (!disk->queue) {
854 			put_disk(disk);
855 			goto out;
856 		}
857 		/*
858 		 * Tell the block layer that we are not a rotational device
859 		 */
860 		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
861 		disk->queue->limits.discard_granularity = 512;
862 		disk->queue->limits.max_discard_sectors = UINT_MAX;
863 		disk->queue->limits.discard_zeroes_data = 0;
864 		blk_queue_max_hw_sectors(disk->queue, 65536);
865 		disk->queue->limits.max_sectors = 256;
866 	}
867 
868 	if (register_blkdev(NBD_MAJOR, "nbd")) {
869 		err = -EIO;
870 		goto out;
871 	}
872 
873 	printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
874 	dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
875 
876 	for (i = 0; i < nbds_max; i++) {
877 		struct gendisk *disk = nbd_dev[i].disk;
878 		nbd_dev[i].file = NULL;
879 		nbd_dev[i].magic = NBD_MAGIC;
880 		nbd_dev[i].flags = 0;
881 		INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
882 		spin_lock_init(&nbd_dev[i].queue_lock);
883 		INIT_LIST_HEAD(&nbd_dev[i].queue_head);
884 		mutex_init(&nbd_dev[i].tx_lock);
885 		init_waitqueue_head(&nbd_dev[i].active_wq);
886 		init_waitqueue_head(&nbd_dev[i].waiting_wq);
887 		nbd_dev[i].blksize = 1024;
888 		nbd_dev[i].bytesize = 0;
889 		disk->major = NBD_MAJOR;
890 		disk->first_minor = i << part_shift;
891 		disk->fops = &nbd_fops;
892 		disk->private_data = &nbd_dev[i];
893 		sprintf(disk->disk_name, "nbd%d", i);
894 		set_capacity(disk, 0);
895 		add_disk(disk);
896 	}
897 
898 	return 0;
899 out:
900 	while (i--) {
901 		blk_cleanup_queue(nbd_dev[i].disk->queue);
902 		put_disk(nbd_dev[i].disk);
903 	}
904 	kfree(nbd_dev);
905 	return err;
906 }
907 
908 static void __exit nbd_cleanup(void)
909 {
910 	int i;
911 	for (i = 0; i < nbds_max; i++) {
912 		struct gendisk *disk = nbd_dev[i].disk;
913 		nbd_dev[i].magic = 0;
914 		if (disk) {
915 			del_gendisk(disk);
916 			blk_cleanup_queue(disk->queue);
917 			put_disk(disk);
918 		}
919 	}
920 	unregister_blkdev(NBD_MAJOR, "nbd");
921 	kfree(nbd_dev);
922 	printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
923 }
924 
925 module_init(nbd_init);
926 module_exit(nbd_cleanup);
927 
928 MODULE_DESCRIPTION("Network Block Device");
929 MODULE_LICENSE("GPL");
930 
931 module_param(nbds_max, int, 0444);
932 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
933 module_param(max_part, int, 0444);
934 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
935 #ifndef NDEBUG
936 module_param(debugflags, int, 0644);
937 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
938 #endif
939