Loading...
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, ¤t->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
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_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
95static 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
108static 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
122static 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
141static 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 */
153static 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, ¤t->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
228static 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 */
240static 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(bvec, 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
296error_out:
297 return -EIO;
298}
299
300static 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
322out:
323 return ERR_PTR(err);
324}
325
326static 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 */
337static 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;
396harderror:
397 nbd->harderror = result;
398 return NULL;
399}
400
401static 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
410static struct device_attribute pid_attr = {
411 .attr = { .name = "pid", .mode = S_IRUGO},
412 .show = pid_show,
413};
414
415static 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
439static 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
474static 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
526error_out:
527 req->errors++;
528 nbd_end_request(req);
529}
530
531static 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
566static 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
604static 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 socket *sock = nbd->sock;
634 nbd->sock = NULL;
635 nbd_clear_que(nbd);
636 BUG_ON(!list_empty(&nbd->queue_head));
637 BUG_ON(!list_empty(&nbd->waiting_queue));
638 kill_bdev(bdev);
639 if (sock)
640 sockfd_put(sock);
641 return 0;
642 }
643
644 case NBD_SET_SOCK: {
645 struct socket *sock;
646 int err;
647 if (nbd->sock)
648 return -EBUSY;
649 sock = sockfd_lookup(arg, &err);
650 if (sock) {
651 nbd->sock = sock;
652 if (max_part > 0)
653 bdev->bd_invalidated = 1;
654 nbd->disconnect = 0; /* we're connected now */
655 return 0;
656 }
657 return -EINVAL;
658 }
659
660 case NBD_SET_BLKSIZE:
661 nbd->blksize = arg;
662 nbd->bytesize &= ~(nbd->blksize-1);
663 bdev->bd_inode->i_size = nbd->bytesize;
664 set_blocksize(bdev, nbd->blksize);
665 set_capacity(nbd->disk, nbd->bytesize >> 9);
666 return 0;
667
668 case NBD_SET_SIZE:
669 nbd->bytesize = arg & ~(nbd->blksize-1);
670 bdev->bd_inode->i_size = nbd->bytesize;
671 set_blocksize(bdev, nbd->blksize);
672 set_capacity(nbd->disk, nbd->bytesize >> 9);
673 return 0;
674
675 case NBD_SET_TIMEOUT:
676 nbd->xmit_timeout = arg * HZ;
677 return 0;
678
679 case NBD_SET_FLAGS:
680 nbd->flags = arg;
681 return 0;
682
683 case NBD_SET_SIZE_BLOCKS:
684 nbd->bytesize = ((u64) arg) * nbd->blksize;
685 bdev->bd_inode->i_size = nbd->bytesize;
686 set_blocksize(bdev, nbd->blksize);
687 set_capacity(nbd->disk, nbd->bytesize >> 9);
688 return 0;
689
690 case NBD_DO_IT: {
691 struct task_struct *thread;
692 struct socket *sock;
693 int error;
694
695 if (nbd->pid)
696 return -EBUSY;
697 if (!nbd->sock)
698 return -EINVAL;
699
700 mutex_unlock(&nbd->tx_lock);
701
702 if (nbd->flags & NBD_FLAG_READ_ONLY)
703 set_device_ro(bdev, true);
704 if (nbd->flags & NBD_FLAG_SEND_TRIM)
705 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
706 nbd->disk->queue);
707 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
708 blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
709 else
710 blk_queue_flush(nbd->disk->queue, 0);
711
712 thread = kthread_create(nbd_thread, nbd, "%s",
713 nbd->disk->disk_name);
714 if (IS_ERR(thread)) {
715 mutex_lock(&nbd->tx_lock);
716 return PTR_ERR(thread);
717 }
718 wake_up_process(thread);
719 error = nbd_do_it(nbd);
720 kthread_stop(thread);
721
722 mutex_lock(&nbd->tx_lock);
723 if (error)
724 return error;
725 sock_shutdown(nbd, 0);
726 sock = nbd->sock;
727 nbd->sock = NULL;
728 nbd_clear_que(nbd);
729 dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
730 kill_bdev(bdev);
731 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
732 set_device_ro(bdev, false);
733 if (sock)
734 sockfd_put(sock);
735 nbd->flags = 0;
736 nbd->bytesize = 0;
737 bdev->bd_inode->i_size = 0;
738 set_capacity(nbd->disk, 0);
739 if (max_part > 0)
740 ioctl_by_bdev(bdev, BLKRRPART, 0);
741 if (nbd->disconnect) /* user requested, ignore socket errors */
742 return 0;
743 return nbd->harderror;
744 }
745
746 case NBD_CLEAR_QUE:
747 /*
748 * This is for compatibility only. The queue is always cleared
749 * by NBD_DO_IT or NBD_CLEAR_SOCK.
750 */
751 return 0;
752
753 case NBD_PRINT_DEBUG:
754 dev_info(disk_to_dev(nbd->disk),
755 "next = %p, prev = %p, head = %p\n",
756 nbd->queue_head.next, nbd->queue_head.prev,
757 &nbd->queue_head);
758 return 0;
759 }
760 return -ENOTTY;
761}
762
763static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
764 unsigned int cmd, unsigned long arg)
765{
766 struct nbd_device *nbd = bdev->bd_disk->private_data;
767 int error;
768
769 if (!capable(CAP_SYS_ADMIN))
770 return -EPERM;
771
772 BUG_ON(nbd->magic != NBD_MAGIC);
773
774 /* Anyone capable of this syscall can do *real bad* things */
775 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
776 nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
777
778 mutex_lock(&nbd->tx_lock);
779 error = __nbd_ioctl(bdev, nbd, cmd, arg);
780 mutex_unlock(&nbd->tx_lock);
781
782 return error;
783}
784
785static const struct block_device_operations nbd_fops =
786{
787 .owner = THIS_MODULE,
788 .ioctl = nbd_ioctl,
789};
790
791/*
792 * And here should be modules and kernel interface
793 * (Just smiley confuses emacs :-)
794 */
795
796static int __init nbd_init(void)
797{
798 int err = -ENOMEM;
799 int i;
800 int part_shift;
801
802 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
803
804 if (max_part < 0) {
805 printk(KERN_ERR "nbd: max_part must be >= 0\n");
806 return -EINVAL;
807 }
808
809 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
810 if (!nbd_dev)
811 return -ENOMEM;
812
813 part_shift = 0;
814 if (max_part > 0) {
815 part_shift = fls(max_part);
816
817 /*
818 * Adjust max_part according to part_shift as it is exported
819 * to user space so that user can know the max number of
820 * partition kernel should be able to manage.
821 *
822 * Note that -1 is required because partition 0 is reserved
823 * for the whole disk.
824 */
825 max_part = (1UL << part_shift) - 1;
826 }
827
828 if ((1UL << part_shift) > DISK_MAX_PARTS)
829 return -EINVAL;
830
831 if (nbds_max > 1UL << (MINORBITS - part_shift))
832 return -EINVAL;
833
834 for (i = 0; i < nbds_max; i++) {
835 struct gendisk *disk = alloc_disk(1 << part_shift);
836 if (!disk)
837 goto out;
838 nbd_dev[i].disk = disk;
839 /*
840 * The new linux 2.5 block layer implementation requires
841 * every gendisk to have its very own request_queue struct.
842 * These structs are big so we dynamically allocate them.
843 */
844 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
845 if (!disk->queue) {
846 put_disk(disk);
847 goto out;
848 }
849 /*
850 * Tell the block layer that we are not a rotational device
851 */
852 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
853 disk->queue->limits.discard_granularity = 512;
854 disk->queue->limits.max_discard_sectors = UINT_MAX;
855 disk->queue->limits.discard_zeroes_data = 0;
856 blk_queue_max_hw_sectors(disk->queue, 65536);
857 disk->queue->limits.max_sectors = 256;
858 }
859
860 if (register_blkdev(NBD_MAJOR, "nbd")) {
861 err = -EIO;
862 goto out;
863 }
864
865 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
866 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
867
868 for (i = 0; i < nbds_max; i++) {
869 struct gendisk *disk = nbd_dev[i].disk;
870 nbd_dev[i].magic = NBD_MAGIC;
871 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
872 spin_lock_init(&nbd_dev[i].queue_lock);
873 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
874 mutex_init(&nbd_dev[i].tx_lock);
875 init_waitqueue_head(&nbd_dev[i].active_wq);
876 init_waitqueue_head(&nbd_dev[i].waiting_wq);
877 nbd_dev[i].blksize = 1024;
878 nbd_dev[i].bytesize = 0;
879 disk->major = NBD_MAJOR;
880 disk->first_minor = i << part_shift;
881 disk->fops = &nbd_fops;
882 disk->private_data = &nbd_dev[i];
883 sprintf(disk->disk_name, "nbd%d", i);
884 set_capacity(disk, 0);
885 add_disk(disk);
886 }
887
888 return 0;
889out:
890 while (i--) {
891 blk_cleanup_queue(nbd_dev[i].disk->queue);
892 put_disk(nbd_dev[i].disk);
893 }
894 kfree(nbd_dev);
895 return err;
896}
897
898static void __exit nbd_cleanup(void)
899{
900 int i;
901 for (i = 0; i < nbds_max; i++) {
902 struct gendisk *disk = nbd_dev[i].disk;
903 nbd_dev[i].magic = 0;
904 if (disk) {
905 del_gendisk(disk);
906 blk_cleanup_queue(disk->queue);
907 put_disk(disk);
908 }
909 }
910 unregister_blkdev(NBD_MAJOR, "nbd");
911 kfree(nbd_dev);
912 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
913}
914
915module_init(nbd_init);
916module_exit(nbd_cleanup);
917
918MODULE_DESCRIPTION("Network Block Device");
919MODULE_LICENSE("GPL");
920
921module_param(nbds_max, int, 0444);
922MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
923module_param(max_part, int, 0444);
924MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
925#ifndef NDEBUG
926module_param(debugflags, int, 0644);
927MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
928#endif