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 * (part of code stolen from loop.c)
12 * 97-3-25 compiled 0-th version, not yet tested it
13 * (it did not work, BTW) (later that day) HEY! it works!
14 * (bit later) hmm, not that much... 2:00am next day:
15 * yes, it works, but it gives something like 50kB/sec
16 * 97-4-01 complete rewrite to make it possible for many requests at
17 * once to be processed
18 * 97-4-11 Making protocol independent of endianity etc.
19 * 97-9-13 Cosmetic changes
20 * 98-5-13 Attempt to make 64-bit-clean on 64-bit machines
21 * 99-1-11 Attempt to make 64-bit-clean on 32-bit machines <ankry@mif.pg.gda.pl>
22 * 01-2-27 Fix to store proper blockcount for kernel (calculated using
23 * BLOCK_SIZE_BITS, not device blocksize) <aga@permonline.ru>
24 * 01-3-11 Make nbd work with new Linux block layer code. It now supports
25 * plugging like all the other block devices. Also added in MSG_MORE to
26 * reduce number of partial TCP segments sent. <steve@chygwyn.com>
27 * 01-12-6 Fix deadlock condition by making queue locks independent of
28 * the transmit lock. <steve@chygwyn.com>
29 * 02-10-11 Allow hung xmit to be aborted via SIGKILL & various fixes.
30 * <Paul.Clements@SteelEye.com> <James.Bottomley@SteelEye.com>
31 * 03-06-22 Make nbd work with new linux 2.5 block layer design. This fixes
32 * memory corruption from module removal and possible memory corruption
33 * from sending/receiving disk data. <ldl@aros.net>
34 * 03-06-23 Cosmetic changes. <ldl@aros.net>
35 * 03-06-23 Enhance diagnostics support. <ldl@aros.net>
36 * 03-06-24 Remove unneeded blksize_bits field from nbd_device struct.
38 * 03-06-24 Cleanup PARANOIA usage & code. <ldl@aros.net>
39 * 04-02-19 Remove PARANOIA, plus various cleanups (Paul Clements)
40 * possible FIXME: make set_sock / set_blksize / set_size / do_it one syscall
41 * why not: would need access_ok and friends, would share yet another
42 * structure with userland
45 #include <linux/major.h>
47 #include <linux/blkdev.h>
48 #include <linux/module.h>
49 #include <linux/init.h>
50 #include <linux/sched.h>
52 #include <linux/bio.h>
53 #include <linux/stat.h>
54 #include <linux/errno.h>
55 #include <linux/file.h>
56 #include <linux/ioctl.h>
59 #include <linux/devfs_fs_kernel.h>
61 #include <asm/uaccess.h>
62 #include <asm/types.h>
64 #include <linux/nbd.h>
66 #define LO_MAGIC 0x68797548
69 #define dprintk(flags, fmt...)
71 #define dprintk(flags, fmt...) do { \
72 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
74 #define DBG_IOCTL 0x0004
75 #define DBG_INIT 0x0010
76 #define DBG_EXIT 0x0020
77 #define DBG_BLKDEV 0x0100
80 static unsigned int debugflags
;
83 static struct nbd_device nbd_dev
[MAX_NBD
];
86 * Use just one lock (or at most 1 per NIC). Two arguments for this:
87 * 1. Each NIC is essentially a synchronization point for all servers
88 * accessed through that NIC so there's no need to have more locks
90 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
91 * down each lock to the point where they're actually slower than just
93 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
95 static DEFINE_SPINLOCK(nbd_lock
);
98 static const char *ioctl_cmd_to_ascii(int cmd
)
101 case NBD_SET_SOCK
: return "set-sock";
102 case NBD_SET_BLKSIZE
: return "set-blksize";
103 case NBD_SET_SIZE
: return "set-size";
104 case NBD_DO_IT
: return "do-it";
105 case NBD_CLEAR_SOCK
: return "clear-sock";
106 case NBD_CLEAR_QUE
: return "clear-que";
107 case NBD_PRINT_DEBUG
: return "print-debug";
108 case NBD_SET_SIZE_BLOCKS
: return "set-size-blocks";
109 case NBD_DISCONNECT
: return "disconnect";
110 case BLKROSET
: return "set-read-only";
111 case BLKFLSBUF
: return "flush-buffer-cache";
116 static const char *nbdcmd_to_ascii(int cmd
)
119 case NBD_CMD_READ
: return "read";
120 case NBD_CMD_WRITE
: return "write";
121 case NBD_CMD_DISC
: return "disconnect";
127 static void nbd_end_request(struct request
*req
)
129 int uptodate
= (req
->errors
== 0) ? 1 : 0;
130 request_queue_t
*q
= req
->q
;
133 dprintk(DBG_BLKDEV
, "%s: request %p: %s\n", req
->rq_disk
->disk_name
,
134 req
, uptodate
? "done": "failed");
136 spin_lock_irqsave(q
->queue_lock
, flags
);
137 if (!end_that_request_first(req
, uptodate
, req
->nr_sectors
)) {
138 end_that_request_last(req
);
140 spin_unlock_irqrestore(q
->queue_lock
, flags
);
144 * Send or receive packet.
146 static int sock_xmit(struct socket
*sock
, int send
, void *buf
, int size
,
155 /* Allow interception of SIGKILL only
156 * Don't allow other signals to interrupt the transmission */
157 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
158 oldset
= current
->blocked
;
159 sigfillset(¤t
->blocked
);
160 sigdelsetmask(¤t
->blocked
, sigmask(SIGKILL
));
162 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
165 sock
->sk
->sk_allocation
= GFP_NOIO
;
170 msg
.msg_control
= NULL
;
171 msg
.msg_controllen
= 0;
173 msg
.msg_flags
= msg_flags
| MSG_NOSIGNAL
;
176 result
= kernel_sendmsg(sock
, &msg
, &iov
, 1, size
);
178 result
= kernel_recvmsg(sock
, &msg
, &iov
, 1, size
, 0);
180 if (signal_pending(current
)) {
182 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
183 printk(KERN_WARNING
"nbd (pid %d: %s) got signal %d\n",
184 current
->pid
, current
->comm
,
185 dequeue_signal(current
, ¤t
->blocked
, &info
));
186 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
193 result
= -EPIPE
; /* short read */
200 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
201 current
->blocked
= oldset
;
203 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
208 static inline int sock_send_bvec(struct socket
*sock
, struct bio_vec
*bvec
,
212 void *kaddr
= kmap(bvec
->bv_page
);
213 result
= sock_xmit(sock
, 1, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
,
215 kunmap(bvec
->bv_page
);
219 static int nbd_send_req(struct nbd_device
*lo
, struct request
*req
)
221 int result
, i
, flags
;
222 struct nbd_request request
;
223 unsigned long size
= req
->nr_sectors
<< 9;
224 struct socket
*sock
= lo
->sock
;
226 request
.magic
= htonl(NBD_REQUEST_MAGIC
);
227 request
.type
= htonl(nbd_cmd(req
));
228 request
.from
= cpu_to_be64((u64
) req
->sector
<< 9);
229 request
.len
= htonl(size
);
230 memcpy(request
.handle
, &req
, sizeof(req
));
234 if (!sock
|| !lo
->sock
) {
235 printk(KERN_ERR
"%s: Attempted send on closed socket\n",
236 lo
->disk
->disk_name
);
240 dprintk(DBG_TX
, "%s: request %p: sending control (%s@%llu,%luB)\n",
241 lo
->disk
->disk_name
, req
,
242 nbdcmd_to_ascii(nbd_cmd(req
)),
243 (unsigned long long)req
->sector
<< 9,
244 req
->nr_sectors
<< 9);
245 result
= sock_xmit(sock
, 1, &request
, sizeof(request
),
246 (nbd_cmd(req
) == NBD_CMD_WRITE
)? MSG_MORE
: 0);
248 printk(KERN_ERR
"%s: Send control failed (result %d)\n",
249 lo
->disk
->disk_name
, result
);
253 if (nbd_cmd(req
) == NBD_CMD_WRITE
) {
256 * we are really probing at internals to determine
257 * whether to set MSG_MORE or not...
259 rq_for_each_bio(bio
, req
) {
260 struct bio_vec
*bvec
;
261 bio_for_each_segment(bvec
, bio
, i
) {
263 if ((i
< (bio
->bi_vcnt
- 1)) || bio
->bi_next
)
265 dprintk(DBG_TX
, "%s: request %p: sending %d bytes data\n",
266 lo
->disk
->disk_name
, req
,
268 result
= sock_send_bvec(sock
, bvec
, flags
);
270 printk(KERN_ERR
"%s: Send data failed (result %d)\n",
286 static struct request
*nbd_find_request(struct nbd_device
*lo
, char *handle
)
289 struct list_head
*tmp
;
290 struct request
*xreq
;
292 memcpy(&xreq
, handle
, sizeof(xreq
));
294 spin_lock(&lo
->queue_lock
);
295 list_for_each(tmp
, &lo
->queue_head
) {
296 req
= list_entry(tmp
, struct request
, queuelist
);
299 list_del_init(&req
->queuelist
);
300 spin_unlock(&lo
->queue_lock
);
303 spin_unlock(&lo
->queue_lock
);
307 static inline int sock_recv_bvec(struct socket
*sock
, struct bio_vec
*bvec
)
310 void *kaddr
= kmap(bvec
->bv_page
);
311 result
= sock_xmit(sock
, 0, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
,
313 kunmap(bvec
->bv_page
);
317 /* NULL returned = something went wrong, inform userspace */
318 static struct request
*nbd_read_stat(struct nbd_device
*lo
)
321 struct nbd_reply reply
;
323 struct socket
*sock
= lo
->sock
;
326 result
= sock_xmit(sock
, 0, &reply
, sizeof(reply
), MSG_WAITALL
);
328 printk(KERN_ERR
"%s: Receive control failed (result %d)\n",
329 lo
->disk
->disk_name
, result
);
332 req
= nbd_find_request(lo
, reply
.handle
);
334 printk(KERN_ERR
"%s: Unexpected reply (%p)\n",
335 lo
->disk
->disk_name
, reply
.handle
);
340 if (ntohl(reply
.magic
) != NBD_REPLY_MAGIC
) {
341 printk(KERN_ERR
"%s: Wrong magic (0x%lx)\n",
343 (unsigned long)ntohl(reply
.magic
));
347 if (ntohl(reply
.error
)) {
348 printk(KERN_ERR
"%s: Other side returned error (%d)\n",
349 lo
->disk
->disk_name
, ntohl(reply
.error
));
354 dprintk(DBG_RX
, "%s: request %p: got reply\n",
355 lo
->disk
->disk_name
, req
);
356 if (nbd_cmd(req
) == NBD_CMD_READ
) {
359 rq_for_each_bio(bio
, req
) {
360 struct bio_vec
*bvec
;
361 bio_for_each_segment(bvec
, bio
, i
) {
362 result
= sock_recv_bvec(sock
, bvec
);
364 printk(KERN_ERR
"%s: Receive data failed (result %d)\n",
369 dprintk(DBG_RX
, "%s: request %p: got %d bytes data\n",
370 lo
->disk
->disk_name
, req
, bvec
->bv_len
);
376 lo
->harderror
= result
;
380 static void nbd_do_it(struct nbd_device
*lo
)
384 BUG_ON(lo
->magic
!= LO_MAGIC
);
386 while ((req
= nbd_read_stat(lo
)) != NULL
)
387 nbd_end_request(req
);
391 static void nbd_clear_que(struct nbd_device
*lo
)
395 BUG_ON(lo
->magic
!= LO_MAGIC
);
399 spin_lock(&lo
->queue_lock
);
400 if (!list_empty(&lo
->queue_head
)) {
401 req
= list_entry(lo
->queue_head
.next
, struct request
, queuelist
);
402 list_del_init(&req
->queuelist
);
404 spin_unlock(&lo
->queue_lock
);
407 nbd_end_request(req
);
413 * We always wait for result of write, for now. It would be nice to make it optional
415 * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
416 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
419 static void do_nbd_request(request_queue_t
* q
)
423 while ((req
= elv_next_request(q
)) != NULL
) {
424 struct nbd_device
*lo
;
426 blkdev_dequeue_request(req
);
427 dprintk(DBG_BLKDEV
, "%s: request %p: dequeued (flags=%lx)\n",
428 req
->rq_disk
->disk_name
, req
, req
->flags
);
430 if (!(req
->flags
& REQ_CMD
))
433 lo
= req
->rq_disk
->private_data
;
435 BUG_ON(lo
->magic
!= LO_MAGIC
);
438 printk(KERN_ERR
"%s: Request when not-ready\n",
439 lo
->disk
->disk_name
);
442 nbd_cmd(req
) = NBD_CMD_READ
;
443 if (rq_data_dir(req
) == WRITE
) {
444 nbd_cmd(req
) = NBD_CMD_WRITE
;
445 if (lo
->flags
& NBD_READ_ONLY
) {
446 printk(KERN_ERR
"%s: Write on read-only\n",
447 lo
->disk
->disk_name
);
453 spin_unlock_irq(q
->queue_lock
);
455 spin_lock(&lo
->queue_lock
);
458 spin_unlock(&lo
->queue_lock
);
459 printk(KERN_ERR
"%s: failed between accept and semaphore, file lost\n",
460 lo
->disk
->disk_name
);
462 nbd_end_request(req
);
463 spin_lock_irq(q
->queue_lock
);
467 list_add(&req
->queuelist
, &lo
->queue_head
);
468 spin_unlock(&lo
->queue_lock
);
470 if (nbd_send_req(lo
, req
) != 0) {
471 printk(KERN_ERR
"%s: Request send failed\n",
472 lo
->disk
->disk_name
);
473 if (nbd_find_request(lo
, (char *)&req
) != NULL
) {
474 /* we still own req */
476 nbd_end_request(req
);
477 } else /* we're racing with nbd_clear_que */
478 printk(KERN_DEBUG
"nbd: can't find req\n");
481 spin_lock_irq(q
->queue_lock
);
486 spin_unlock(q
->queue_lock
);
487 nbd_end_request(req
);
488 spin_lock(q
->queue_lock
);
493 static int nbd_ioctl(struct inode
*inode
, struct file
*file
,
494 unsigned int cmd
, unsigned long arg
)
496 struct nbd_device
*lo
= inode
->i_bdev
->bd_disk
->private_data
;
498 struct request sreq
;
500 if (!capable(CAP_SYS_ADMIN
))
503 BUG_ON(lo
->magic
!= LO_MAGIC
);
505 /* Anyone capable of this syscall can do *real bad* things */
506 dprintk(DBG_IOCTL
, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
507 lo
->disk
->disk_name
, ioctl_cmd_to_ascii(cmd
), cmd
, arg
);
511 printk(KERN_INFO
"%s: NBD_DISCONNECT\n", lo
->disk
->disk_name
);
512 sreq
.flags
= REQ_SPECIAL
;
513 nbd_cmd(&sreq
) = NBD_CMD_DISC
;
515 * Set these to sane values in case server implementation
516 * fails to check the request type first and also to keep
517 * debugging output cleaner.
523 nbd_send_req(lo
, &sreq
);
531 spin_lock(&lo
->queue_lock
);
534 spin_unlock(&lo
->queue_lock
);
536 spin_lock(&lo
->queue_lock
);
537 if (!list_empty(&lo
->queue_head
)) {
538 printk(KERN_ERR
"nbd: disconnect: some requests are in progress -> please try again.\n");
541 spin_unlock(&lo
->queue_lock
);
551 inode
= file
->f_dentry
->d_inode
;
552 if (S_ISSOCK(inode
->i_mode
)) {
554 lo
->sock
= SOCKET_I(inode
);
561 case NBD_SET_BLKSIZE
:
563 lo
->bytesize
&= ~(lo
->blksize
-1);
564 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
565 set_blocksize(inode
->i_bdev
, lo
->blksize
);
566 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
569 lo
->bytesize
= arg
& ~(lo
->blksize
-1);
570 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
571 set_blocksize(inode
->i_bdev
, lo
->blksize
);
572 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
574 case NBD_SET_SIZE_BLOCKS
:
575 lo
->bytesize
= ((u64
) arg
) * lo
->blksize
;
576 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
577 set_blocksize(inode
->i_bdev
, lo
->blksize
);
578 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
584 /* on return tidy up in case we have a signal */
585 /* Forcibly shutdown the socket causing all listeners
588 * FIXME: This code is duplicated from sys_shutdown, but
589 * there should be a more generic interface rather than
590 * calling socket ops directly here */
593 printk(KERN_WARNING
"%s: shutting down socket\n",
594 lo
->disk
->disk_name
);
595 lo
->sock
->ops
->shutdown(lo
->sock
,
596 SEND_SHUTDOWN
|RCV_SHUTDOWN
);
600 spin_lock(&lo
->queue_lock
);
603 spin_unlock(&lo
->queue_lock
);
605 printk(KERN_WARNING
"%s: queue cleared\n", lo
->disk
->disk_name
);
608 return lo
->harderror
;
613 return 0; /* probably should be error, but that would
614 * break "nbd-client -d", so just return 0 */
619 case NBD_PRINT_DEBUG
:
620 printk(KERN_INFO
"%s: next = %p, prev = %p, head = %p\n",
621 inode
->i_bdev
->bd_disk
->disk_name
,
622 lo
->queue_head
.next
, lo
->queue_head
.prev
,
629 static struct block_device_operations nbd_fops
=
631 .owner
= THIS_MODULE
,
636 * And here should be modules and kernel interface
637 * (Just smiley confuses emacs :-)
640 static int __init
nbd_init(void)
645 if (sizeof(struct nbd_request
) != 28) {
646 printk(KERN_CRIT
"nbd: sizeof nbd_request needs to be 28 in order to work!\n" );
650 for (i
= 0; i
< MAX_NBD
; i
++) {
651 struct gendisk
*disk
= alloc_disk(1);
654 nbd_dev
[i
].disk
= disk
;
656 * The new linux 2.5 block layer implementation requires
657 * every gendisk to have its very own request_queue struct.
658 * These structs are big so we dynamically allocate them.
660 disk
->queue
= blk_init_queue(do_nbd_request
, &nbd_lock
);
667 if (register_blkdev(NBD_MAJOR
, "nbd")) {
672 printk(KERN_INFO
"nbd: registered device at major %d\n", NBD_MAJOR
);
673 dprintk(DBG_INIT
, "nbd: debugflags=0x%x\n", debugflags
);
676 for (i
= 0; i
< MAX_NBD
; i
++) {
677 struct gendisk
*disk
= nbd_dev
[i
].disk
;
678 nbd_dev
[i
].file
= NULL
;
679 nbd_dev
[i
].magic
= LO_MAGIC
;
680 nbd_dev
[i
].flags
= 0;
681 spin_lock_init(&nbd_dev
[i
].queue_lock
);
682 INIT_LIST_HEAD(&nbd_dev
[i
].queue_head
);
683 init_MUTEX(&nbd_dev
[i
].tx_lock
);
684 nbd_dev
[i
].blksize
= 1024;
685 nbd_dev
[i
].bytesize
= 0x7ffffc00ULL
<< 10; /* 2TB */
686 disk
->major
= NBD_MAJOR
;
687 disk
->first_minor
= i
;
688 disk
->fops
= &nbd_fops
;
689 disk
->private_data
= &nbd_dev
[i
];
690 disk
->flags
|= GENHD_FL_SUPPRESS_PARTITION_INFO
;
691 sprintf(disk
->disk_name
, "nbd%d", i
);
692 sprintf(disk
->devfs_name
, "nbd/%d", i
);
693 set_capacity(disk
, 0x7ffffc00ULL
<< 1); /* 2 TB */
700 blk_cleanup_queue(nbd_dev
[i
].disk
->queue
);
701 put_disk(nbd_dev
[i
].disk
);
706 static void __exit
nbd_cleanup(void)
709 for (i
= 0; i
< MAX_NBD
; i
++) {
710 struct gendisk
*disk
= nbd_dev
[i
].disk
;
713 blk_cleanup_queue(disk
->queue
);
718 unregister_blkdev(NBD_MAJOR
, "nbd");
719 printk(KERN_INFO
"nbd: unregistered device at major %d\n", NBD_MAJOR
);
722 module_init(nbd_init
);
723 module_exit(nbd_cleanup
);
725 MODULE_DESCRIPTION("Network Block Device");
726 MODULE_LICENSE("GPL");
729 module_param(debugflags
, int, 0644);
730 MODULE_PARM_DESC(debugflags
, "flags for controlling debug output");