2 * Network block device - make block devices work over TCP
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.
7 * Copyright 1997-2000 Pavel Machek <pavel@ucw.cz>
8 * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
10 * This file is released under GPLv2 or later.
12 * (part of code stolen from loop.c)
15 #include <linux/major.h>
17 #include <linux/blkdev.h>
18 #include <linux/module.h>
19 #include <linux/init.h>
20 #include <linux/sched.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/compiler.h>
28 #include <linux/err.h>
29 #include <linux/kernel.h>
32 #include <asm/uaccess.h>
33 #include <asm/system.h>
34 #include <asm/types.h>
36 #include <linux/nbd.h>
38 #define LO_MAGIC 0x68797548
41 #define dprintk(flags, fmt...)
43 #define dprintk(flags, fmt...) do { \
44 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
46 #define DBG_IOCTL 0x0004
47 #define DBG_INIT 0x0010
48 #define DBG_EXIT 0x0020
49 #define DBG_BLKDEV 0x0100
52 static unsigned int debugflags
;
55 static unsigned int nbds_max
= 16;
56 static struct nbd_device nbd_dev
[MAX_NBD
];
59 * Use just one lock (or at most 1 per NIC). Two arguments for this:
60 * 1. Each NIC is essentially a synchronization point for all servers
61 * accessed through that NIC so there's no need to have more locks
63 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
64 * down each lock to the point where they're actually slower than just
66 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
68 static DEFINE_SPINLOCK(nbd_lock
);
71 static const char *ioctl_cmd_to_ascii(int cmd
)
74 case NBD_SET_SOCK
: return "set-sock";
75 case NBD_SET_BLKSIZE
: return "set-blksize";
76 case NBD_SET_SIZE
: return "set-size";
77 case NBD_DO_IT
: return "do-it";
78 case NBD_CLEAR_SOCK
: return "clear-sock";
79 case NBD_CLEAR_QUE
: return "clear-que";
80 case NBD_PRINT_DEBUG
: return "print-debug";
81 case NBD_SET_SIZE_BLOCKS
: return "set-size-blocks";
82 case NBD_DISCONNECT
: return "disconnect";
83 case BLKROSET
: return "set-read-only";
84 case BLKFLSBUF
: return "flush-buffer-cache";
89 static const char *nbdcmd_to_ascii(int cmd
)
92 case NBD_CMD_READ
: return "read";
93 case NBD_CMD_WRITE
: return "write";
94 case NBD_CMD_DISC
: return "disconnect";
100 static void nbd_end_request(struct request
*req
)
102 int uptodate
= (req
->errors
== 0) ? 1 : 0;
103 request_queue_t
*q
= req
->q
;
106 dprintk(DBG_BLKDEV
, "%s: request %p: %s\n", req
->rq_disk
->disk_name
,
107 req
, uptodate
? "done": "failed");
109 spin_lock_irqsave(q
->queue_lock
, flags
);
110 if (!end_that_request_first(req
, uptodate
, req
->nr_sectors
)) {
111 end_that_request_last(req
, uptodate
);
113 spin_unlock_irqrestore(q
->queue_lock
, flags
);
117 * Send or receive packet.
119 static int sock_xmit(struct socket
*sock
, int send
, void *buf
, int size
,
128 /* Allow interception of SIGKILL only
129 * Don't allow other signals to interrupt the transmission */
130 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
131 oldset
= current
->blocked
;
132 sigfillset(¤t
->blocked
);
133 sigdelsetmask(¤t
->blocked
, sigmask(SIGKILL
));
135 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
138 sock
->sk
->sk_allocation
= GFP_NOIO
;
143 msg
.msg_control
= NULL
;
144 msg
.msg_controllen
= 0;
145 msg
.msg_flags
= msg_flags
| MSG_NOSIGNAL
;
148 result
= kernel_sendmsg(sock
, &msg
, &iov
, 1, size
);
150 result
= kernel_recvmsg(sock
, &msg
, &iov
, 1, size
, 0);
152 if (signal_pending(current
)) {
154 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
155 printk(KERN_WARNING
"nbd (pid %d: %s) got signal %d\n",
156 current
->pid
, current
->comm
,
157 dequeue_signal(current
, ¤t
->blocked
, &info
));
158 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
165 result
= -EPIPE
; /* short read */
172 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
173 current
->blocked
= oldset
;
175 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
180 static inline int sock_send_bvec(struct socket
*sock
, struct bio_vec
*bvec
,
184 void *kaddr
= kmap(bvec
->bv_page
);
185 result
= sock_xmit(sock
, 1, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
,
187 kunmap(bvec
->bv_page
);
191 static int nbd_send_req(struct nbd_device
*lo
, struct request
*req
)
193 int result
, i
, flags
;
194 struct nbd_request request
;
195 unsigned long size
= req
->nr_sectors
<< 9;
196 struct socket
*sock
= lo
->sock
;
198 request
.magic
= htonl(NBD_REQUEST_MAGIC
);
199 request
.type
= htonl(nbd_cmd(req
));
200 request
.from
= cpu_to_be64((u64
) req
->sector
<< 9);
201 request
.len
= htonl(size
);
202 memcpy(request
.handle
, &req
, sizeof(req
));
204 dprintk(DBG_TX
, "%s: request %p: sending control (%s@%llu,%luB)\n",
205 lo
->disk
->disk_name
, req
,
206 nbdcmd_to_ascii(nbd_cmd(req
)),
207 (unsigned long long)req
->sector
<< 9,
208 req
->nr_sectors
<< 9);
209 result
= sock_xmit(sock
, 1, &request
, sizeof(request
),
210 (nbd_cmd(req
) == NBD_CMD_WRITE
)? MSG_MORE
: 0);
212 printk(KERN_ERR
"%s: Send control failed (result %d)\n",
213 lo
->disk
->disk_name
, result
);
217 if (nbd_cmd(req
) == NBD_CMD_WRITE
) {
220 * we are really probing at internals to determine
221 * whether to set MSG_MORE or not...
223 rq_for_each_bio(bio
, req
) {
224 struct bio_vec
*bvec
;
225 bio_for_each_segment(bvec
, bio
, i
) {
227 if ((i
< (bio
->bi_vcnt
- 1)) || bio
->bi_next
)
229 dprintk(DBG_TX
, "%s: request %p: sending %d bytes data\n",
230 lo
->disk
->disk_name
, req
,
232 result
= sock_send_bvec(sock
, bvec
, flags
);
234 printk(KERN_ERR
"%s: Send data failed (result %d)\n",
248 static struct request
*nbd_find_request(struct nbd_device
*lo
, char *handle
)
251 struct list_head
*tmp
;
252 struct request
*xreq
;
255 memcpy(&xreq
, handle
, sizeof(xreq
));
257 err
= wait_event_interruptible(lo
->active_wq
, lo
->active_req
!= xreq
);
261 spin_lock(&lo
->queue_lock
);
262 list_for_each(tmp
, &lo
->queue_head
) {
263 req
= list_entry(tmp
, struct request
, queuelist
);
266 list_del_init(&req
->queuelist
);
267 spin_unlock(&lo
->queue_lock
);
270 spin_unlock(&lo
->queue_lock
);
278 static inline int sock_recv_bvec(struct socket
*sock
, struct bio_vec
*bvec
)
281 void *kaddr
= kmap(bvec
->bv_page
);
282 result
= sock_xmit(sock
, 0, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
,
284 kunmap(bvec
->bv_page
);
288 /* NULL returned = something went wrong, inform userspace */
289 static struct request
*nbd_read_stat(struct nbd_device
*lo
)
292 struct nbd_reply reply
;
294 struct socket
*sock
= lo
->sock
;
297 result
= sock_xmit(sock
, 0, &reply
, sizeof(reply
), MSG_WAITALL
);
299 printk(KERN_ERR
"%s: Receive control failed (result %d)\n",
300 lo
->disk
->disk_name
, result
);
304 if (ntohl(reply
.magic
) != NBD_REPLY_MAGIC
) {
305 printk(KERN_ERR
"%s: Wrong magic (0x%lx)\n",
307 (unsigned long)ntohl(reply
.magic
));
312 req
= nbd_find_request(lo
, reply
.handle
);
313 if (unlikely(IS_ERR(req
))) {
314 result
= PTR_ERR(req
);
315 if (result
!= -ENOENT
)
318 printk(KERN_ERR
"%s: Unexpected reply (%p)\n",
319 lo
->disk
->disk_name
, reply
.handle
);
324 if (ntohl(reply
.error
)) {
325 printk(KERN_ERR
"%s: Other side returned error (%d)\n",
326 lo
->disk
->disk_name
, ntohl(reply
.error
));
331 dprintk(DBG_RX
, "%s: request %p: got reply\n",
332 lo
->disk
->disk_name
, req
);
333 if (nbd_cmd(req
) == NBD_CMD_READ
) {
336 rq_for_each_bio(bio
, req
) {
337 struct bio_vec
*bvec
;
338 bio_for_each_segment(bvec
, bio
, i
) {
339 result
= sock_recv_bvec(sock
, bvec
);
341 printk(KERN_ERR
"%s: Receive data failed (result %d)\n",
347 dprintk(DBG_RX
, "%s: request %p: got %d bytes data\n",
348 lo
->disk
->disk_name
, req
, bvec
->bv_len
);
354 lo
->harderror
= result
;
358 static void nbd_do_it(struct nbd_device
*lo
)
362 BUG_ON(lo
->magic
!= LO_MAGIC
);
364 while ((req
= nbd_read_stat(lo
)) != NULL
)
365 nbd_end_request(req
);
369 static void nbd_clear_que(struct nbd_device
*lo
)
373 BUG_ON(lo
->magic
!= LO_MAGIC
);
376 * Because we have set lo->sock to NULL under the tx_lock, all
377 * modifications to the list must have completed by now. For
378 * the same reason, the active_req must be NULL.
380 * As a consequence, we don't need to take the spin lock while
381 * purging the list here.
384 BUG_ON(lo
->active_req
);
386 while (!list_empty(&lo
->queue_head
)) {
387 req
= list_entry(lo
->queue_head
.next
, struct request
,
389 list_del_init(&req
->queuelist
);
391 nbd_end_request(req
);
396 * We always wait for result of write, for now. It would be nice to make it optional
398 * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
399 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
402 static void do_nbd_request(request_queue_t
* q
)
406 while ((req
= elv_next_request(q
)) != NULL
) {
407 struct nbd_device
*lo
;
409 blkdev_dequeue_request(req
);
410 dprintk(DBG_BLKDEV
, "%s: request %p: dequeued (flags=%lx)\n",
411 req
->rq_disk
->disk_name
, req
, req
->flags
);
413 if (!(req
->flags
& REQ_CMD
))
416 lo
= req
->rq_disk
->private_data
;
418 BUG_ON(lo
->magic
!= LO_MAGIC
);
420 nbd_cmd(req
) = NBD_CMD_READ
;
421 if (rq_data_dir(req
) == WRITE
) {
422 nbd_cmd(req
) = NBD_CMD_WRITE
;
423 if (lo
->flags
& NBD_READ_ONLY
) {
424 printk(KERN_ERR
"%s: Write on read-only\n",
425 lo
->disk
->disk_name
);
431 spin_unlock_irq(q
->queue_lock
);
433 mutex_lock(&lo
->tx_lock
);
434 if (unlikely(!lo
->sock
)) {
435 mutex_unlock(&lo
->tx_lock
);
436 printk(KERN_ERR
"%s: Attempted send on closed socket\n",
437 lo
->disk
->disk_name
);
439 nbd_end_request(req
);
440 spin_lock_irq(q
->queue_lock
);
444 lo
->active_req
= req
;
446 if (nbd_send_req(lo
, req
) != 0) {
447 printk(KERN_ERR
"%s: Request send failed\n",
448 lo
->disk
->disk_name
);
450 nbd_end_request(req
);
452 spin_lock(&lo
->queue_lock
);
453 list_add(&req
->queuelist
, &lo
->queue_head
);
454 spin_unlock(&lo
->queue_lock
);
457 lo
->active_req
= NULL
;
458 mutex_unlock(&lo
->tx_lock
);
459 wake_up_all(&lo
->active_wq
);
461 spin_lock_irq(q
->queue_lock
);
466 spin_unlock(q
->queue_lock
);
467 nbd_end_request(req
);
468 spin_lock(q
->queue_lock
);
473 static int nbd_ioctl(struct inode
*inode
, struct file
*file
,
474 unsigned int cmd
, unsigned long arg
)
476 struct nbd_device
*lo
= inode
->i_bdev
->bd_disk
->private_data
;
478 struct request sreq
;
480 if (!capable(CAP_SYS_ADMIN
))
483 BUG_ON(lo
->magic
!= LO_MAGIC
);
485 /* Anyone capable of this syscall can do *real bad* things */
486 dprintk(DBG_IOCTL
, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
487 lo
->disk
->disk_name
, ioctl_cmd_to_ascii(cmd
), cmd
, arg
);
491 printk(KERN_INFO
"%s: NBD_DISCONNECT\n", lo
->disk
->disk_name
);
492 sreq
.flags
= REQ_SPECIAL
;
493 nbd_cmd(&sreq
) = NBD_CMD_DISC
;
495 * Set these to sane values in case server implementation
496 * fails to check the request type first and also to keep
497 * debugging output cleaner.
503 nbd_send_req(lo
, &sreq
);
508 mutex_lock(&lo
->tx_lock
);
510 mutex_unlock(&lo
->tx_lock
);
514 BUG_ON(!list_empty(&lo
->queue_head
));
524 inode
= file
->f_dentry
->d_inode
;
525 if (S_ISSOCK(inode
->i_mode
)) {
527 lo
->sock
= SOCKET_I(inode
);
534 case NBD_SET_BLKSIZE
:
536 lo
->bytesize
&= ~(lo
->blksize
-1);
537 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
538 set_blocksize(inode
->i_bdev
, lo
->blksize
);
539 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
542 lo
->bytesize
= arg
& ~(lo
->blksize
-1);
543 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
544 set_blocksize(inode
->i_bdev
, lo
->blksize
);
545 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
547 case NBD_SET_SIZE_BLOCKS
:
548 lo
->bytesize
= ((u64
) arg
) * lo
->blksize
;
549 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
550 set_blocksize(inode
->i_bdev
, lo
->blksize
);
551 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
557 /* on return tidy up in case we have a signal */
558 /* Forcibly shutdown the socket causing all listeners
561 * FIXME: This code is duplicated from sys_shutdown, but
562 * there should be a more generic interface rather than
563 * calling socket ops directly here */
564 mutex_lock(&lo
->tx_lock
);
566 printk(KERN_WARNING
"%s: shutting down socket\n",
567 lo
->disk
->disk_name
);
568 lo
->sock
->ops
->shutdown(lo
->sock
,
569 SEND_SHUTDOWN
|RCV_SHUTDOWN
);
572 mutex_unlock(&lo
->tx_lock
);
576 printk(KERN_WARNING
"%s: queue cleared\n", lo
->disk
->disk_name
);
579 return lo
->harderror
;
582 * This is for compatibility only. The queue is always cleared
583 * by NBD_DO_IT or NBD_CLEAR_SOCK.
585 BUG_ON(!lo
->sock
&& !list_empty(&lo
->queue_head
));
587 case NBD_PRINT_DEBUG
:
588 printk(KERN_INFO
"%s: next = %p, prev = %p, head = %p\n",
589 inode
->i_bdev
->bd_disk
->disk_name
,
590 lo
->queue_head
.next
, lo
->queue_head
.prev
,
597 static struct block_device_operations nbd_fops
=
599 .owner
= THIS_MODULE
,
604 * And here should be modules and kernel interface
605 * (Just smiley confuses emacs :-)
608 static int __init
nbd_init(void)
613 BUILD_BUG_ON(sizeof(struct nbd_request
) != 28);
615 if (nbds_max
> MAX_NBD
) {
616 printk(KERN_CRIT
"nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD
,
621 for (i
= 0; i
< nbds_max
; i
++) {
622 struct gendisk
*disk
= alloc_disk(1);
625 nbd_dev
[i
].disk
= disk
;
627 * The new linux 2.5 block layer implementation requires
628 * every gendisk to have its very own request_queue struct.
629 * These structs are big so we dynamically allocate them.
631 disk
->queue
= blk_init_queue(do_nbd_request
, &nbd_lock
);
638 if (register_blkdev(NBD_MAJOR
, "nbd")) {
643 printk(KERN_INFO
"nbd: registered device at major %d\n", NBD_MAJOR
);
644 dprintk(DBG_INIT
, "nbd: debugflags=0x%x\n", debugflags
);
646 for (i
= 0; i
< nbds_max
; i
++) {
647 struct gendisk
*disk
= nbd_dev
[i
].disk
;
648 nbd_dev
[i
].file
= NULL
;
649 nbd_dev
[i
].magic
= LO_MAGIC
;
650 nbd_dev
[i
].flags
= 0;
651 spin_lock_init(&nbd_dev
[i
].queue_lock
);
652 INIT_LIST_HEAD(&nbd_dev
[i
].queue_head
);
653 mutex_init(&nbd_dev
[i
].tx_lock
);
654 init_waitqueue_head(&nbd_dev
[i
].active_wq
);
655 nbd_dev
[i
].blksize
= 1024;
656 nbd_dev
[i
].bytesize
= 0x7ffffc00ULL
<< 10; /* 2TB */
657 disk
->major
= NBD_MAJOR
;
658 disk
->first_minor
= i
;
659 disk
->fops
= &nbd_fops
;
660 disk
->private_data
= &nbd_dev
[i
];
661 disk
->flags
|= GENHD_FL_SUPPRESS_PARTITION_INFO
;
662 sprintf(disk
->disk_name
, "nbd%d", i
);
663 set_capacity(disk
, 0x7ffffc00ULL
<< 1); /* 2 TB */
670 blk_cleanup_queue(nbd_dev
[i
].disk
->queue
);
671 put_disk(nbd_dev
[i
].disk
);
676 static void __exit
nbd_cleanup(void)
679 for (i
= 0; i
< nbds_max
; i
++) {
680 struct gendisk
*disk
= nbd_dev
[i
].disk
;
681 nbd_dev
[i
].magic
= 0;
684 blk_cleanup_queue(disk
->queue
);
688 unregister_blkdev(NBD_MAJOR
, "nbd");
689 printk(KERN_INFO
"nbd: unregistered device at major %d\n", NBD_MAJOR
);
692 module_init(nbd_init
);
693 module_exit(nbd_cleanup
);
695 MODULE_DESCRIPTION("Network Block Device");
696 MODULE_LICENSE("GPL");
698 module_param(nbds_max
, int, 0444);
699 MODULE_PARM_DESC(nbds_max
, "How many network block devices to initialize.");
701 module_param(debugflags
, int, 0644);
702 MODULE_PARM_DESC(debugflags
, "flags for controlling debug output");