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
;
81 static unsigned int nbds_max
= 16;
84 static struct nbd_device nbd_dev
[MAX_NBD
];
87 * Use just one lock (or at most 1 per NIC). Two arguments for this:
88 * 1. Each NIC is essentially a synchronization point for all servers
89 * accessed through that NIC so there's no need to have more locks
91 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
92 * down each lock to the point where they're actually slower than just
94 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
96 static DEFINE_SPINLOCK(nbd_lock
);
99 static const char *ioctl_cmd_to_ascii(int cmd
)
102 case NBD_SET_SOCK
: return "set-sock";
103 case NBD_SET_BLKSIZE
: return "set-blksize";
104 case NBD_SET_SIZE
: return "set-size";
105 case NBD_DO_IT
: return "do-it";
106 case NBD_CLEAR_SOCK
: return "clear-sock";
107 case NBD_CLEAR_QUE
: return "clear-que";
108 case NBD_PRINT_DEBUG
: return "print-debug";
109 case NBD_SET_SIZE_BLOCKS
: return "set-size-blocks";
110 case NBD_DISCONNECT
: return "disconnect";
111 case BLKROSET
: return "set-read-only";
112 case BLKFLSBUF
: return "flush-buffer-cache";
117 static const char *nbdcmd_to_ascii(int cmd
)
120 case NBD_CMD_READ
: return "read";
121 case NBD_CMD_WRITE
: return "write";
122 case NBD_CMD_DISC
: return "disconnect";
128 static void nbd_end_request(struct request
*req
)
130 int uptodate
= (req
->errors
== 0) ? 1 : 0;
131 request_queue_t
*q
= req
->q
;
134 dprintk(DBG_BLKDEV
, "%s: request %p: %s\n", req
->rq_disk
->disk_name
,
135 req
, uptodate
? "done": "failed");
137 spin_lock_irqsave(q
->queue_lock
, flags
);
138 if (!end_that_request_first(req
, uptodate
, req
->nr_sectors
)) {
139 end_that_request_last(req
);
141 spin_unlock_irqrestore(q
->queue_lock
, flags
);
145 * Send or receive packet.
147 static int sock_xmit(struct socket
*sock
, int send
, void *buf
, int size
,
156 /* Allow interception of SIGKILL only
157 * Don't allow other signals to interrupt the transmission */
158 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
159 oldset
= current
->blocked
;
160 sigfillset(¤t
->blocked
);
161 sigdelsetmask(¤t
->blocked
, sigmask(SIGKILL
));
163 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
166 sock
->sk
->sk_allocation
= GFP_NOIO
;
171 msg
.msg_control
= NULL
;
172 msg
.msg_controllen
= 0;
174 msg
.msg_flags
= msg_flags
| MSG_NOSIGNAL
;
177 result
= kernel_sendmsg(sock
, &msg
, &iov
, 1, size
);
179 result
= kernel_recvmsg(sock
, &msg
, &iov
, 1, size
, 0);
181 if (signal_pending(current
)) {
183 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
184 printk(KERN_WARNING
"nbd (pid %d: %s) got signal %d\n",
185 current
->pid
, current
->comm
,
186 dequeue_signal(current
, ¤t
->blocked
, &info
));
187 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
194 result
= -EPIPE
; /* short read */
201 spin_lock_irqsave(¤t
->sighand
->siglock
, flags
);
202 current
->blocked
= oldset
;
204 spin_unlock_irqrestore(¤t
->sighand
->siglock
, flags
);
209 static inline int sock_send_bvec(struct socket
*sock
, struct bio_vec
*bvec
,
213 void *kaddr
= kmap(bvec
->bv_page
);
214 result
= sock_xmit(sock
, 1, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
,
216 kunmap(bvec
->bv_page
);
220 static int nbd_send_req(struct nbd_device
*lo
, struct request
*req
)
222 int result
, i
, flags
;
223 struct nbd_request request
;
224 unsigned long size
= req
->nr_sectors
<< 9;
225 struct socket
*sock
= lo
->sock
;
227 request
.magic
= htonl(NBD_REQUEST_MAGIC
);
228 request
.type
= htonl(nbd_cmd(req
));
229 request
.from
= cpu_to_be64((u64
) req
->sector
<< 9);
230 request
.len
= htonl(size
);
231 memcpy(request
.handle
, &req
, sizeof(req
));
235 if (!sock
|| !lo
->sock
) {
236 printk(KERN_ERR
"%s: Attempted send on closed socket\n",
237 lo
->disk
->disk_name
);
241 dprintk(DBG_TX
, "%s: request %p: sending control (%s@%llu,%luB)\n",
242 lo
->disk
->disk_name
, req
,
243 nbdcmd_to_ascii(nbd_cmd(req
)),
244 (unsigned long long)req
->sector
<< 9,
245 req
->nr_sectors
<< 9);
246 result
= sock_xmit(sock
, 1, &request
, sizeof(request
),
247 (nbd_cmd(req
) == NBD_CMD_WRITE
)? MSG_MORE
: 0);
249 printk(KERN_ERR
"%s: Send control failed (result %d)\n",
250 lo
->disk
->disk_name
, result
);
254 if (nbd_cmd(req
) == NBD_CMD_WRITE
) {
257 * we are really probing at internals to determine
258 * whether to set MSG_MORE or not...
260 rq_for_each_bio(bio
, req
) {
261 struct bio_vec
*bvec
;
262 bio_for_each_segment(bvec
, bio
, i
) {
264 if ((i
< (bio
->bi_vcnt
- 1)) || bio
->bi_next
)
266 dprintk(DBG_TX
, "%s: request %p: sending %d bytes data\n",
267 lo
->disk
->disk_name
, req
,
269 result
= sock_send_bvec(sock
, bvec
, flags
);
271 printk(KERN_ERR
"%s: Send data failed (result %d)\n",
287 static struct request
*nbd_find_request(struct nbd_device
*lo
, char *handle
)
290 struct list_head
*tmp
;
291 struct request
*xreq
;
293 memcpy(&xreq
, handle
, sizeof(xreq
));
295 spin_lock(&lo
->queue_lock
);
296 list_for_each(tmp
, &lo
->queue_head
) {
297 req
= list_entry(tmp
, struct request
, queuelist
);
300 list_del_init(&req
->queuelist
);
301 spin_unlock(&lo
->queue_lock
);
304 spin_unlock(&lo
->queue_lock
);
308 static inline int sock_recv_bvec(struct socket
*sock
, struct bio_vec
*bvec
)
311 void *kaddr
= kmap(bvec
->bv_page
);
312 result
= sock_xmit(sock
, 0, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
,
314 kunmap(bvec
->bv_page
);
318 /* NULL returned = something went wrong, inform userspace */
319 static struct request
*nbd_read_stat(struct nbd_device
*lo
)
322 struct nbd_reply reply
;
324 struct socket
*sock
= lo
->sock
;
327 result
= sock_xmit(sock
, 0, &reply
, sizeof(reply
), MSG_WAITALL
);
329 printk(KERN_ERR
"%s: Receive control failed (result %d)\n",
330 lo
->disk
->disk_name
, result
);
333 req
= nbd_find_request(lo
, reply
.handle
);
335 printk(KERN_ERR
"%s: Unexpected reply (%p)\n",
336 lo
->disk
->disk_name
, reply
.handle
);
341 if (ntohl(reply
.magic
) != NBD_REPLY_MAGIC
) {
342 printk(KERN_ERR
"%s: Wrong magic (0x%lx)\n",
344 (unsigned long)ntohl(reply
.magic
));
348 if (ntohl(reply
.error
)) {
349 printk(KERN_ERR
"%s: Other side returned error (%d)\n",
350 lo
->disk
->disk_name
, ntohl(reply
.error
));
355 dprintk(DBG_RX
, "%s: request %p: got reply\n",
356 lo
->disk
->disk_name
, req
);
357 if (nbd_cmd(req
) == NBD_CMD_READ
) {
360 rq_for_each_bio(bio
, req
) {
361 struct bio_vec
*bvec
;
362 bio_for_each_segment(bvec
, bio
, i
) {
363 result
= sock_recv_bvec(sock
, bvec
);
365 printk(KERN_ERR
"%s: Receive data failed (result %d)\n",
370 dprintk(DBG_RX
, "%s: request %p: got %d bytes data\n",
371 lo
->disk
->disk_name
, req
, bvec
->bv_len
);
377 lo
->harderror
= result
;
381 static void nbd_do_it(struct nbd_device
*lo
)
385 BUG_ON(lo
->magic
!= LO_MAGIC
);
387 while ((req
= nbd_read_stat(lo
)) != NULL
)
388 nbd_end_request(req
);
392 static void nbd_clear_que(struct nbd_device
*lo
)
396 BUG_ON(lo
->magic
!= LO_MAGIC
);
400 spin_lock(&lo
->queue_lock
);
401 if (!list_empty(&lo
->queue_head
)) {
402 req
= list_entry(lo
->queue_head
.next
, struct request
, queuelist
);
403 list_del_init(&req
->queuelist
);
405 spin_unlock(&lo
->queue_lock
);
408 nbd_end_request(req
);
414 * We always wait for result of write, for now. It would be nice to make it optional
416 * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
417 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
420 static void do_nbd_request(request_queue_t
* q
)
424 while ((req
= elv_next_request(q
)) != NULL
) {
425 struct nbd_device
*lo
;
427 blkdev_dequeue_request(req
);
428 dprintk(DBG_BLKDEV
, "%s: request %p: dequeued (flags=%lx)\n",
429 req
->rq_disk
->disk_name
, req
, req
->flags
);
431 if (!(req
->flags
& REQ_CMD
))
434 lo
= req
->rq_disk
->private_data
;
436 BUG_ON(lo
->magic
!= LO_MAGIC
);
439 printk(KERN_ERR
"%s: Request when not-ready\n",
440 lo
->disk
->disk_name
);
443 nbd_cmd(req
) = NBD_CMD_READ
;
444 if (rq_data_dir(req
) == WRITE
) {
445 nbd_cmd(req
) = NBD_CMD_WRITE
;
446 if (lo
->flags
& NBD_READ_ONLY
) {
447 printk(KERN_ERR
"%s: Write on read-only\n",
448 lo
->disk
->disk_name
);
454 spin_unlock_irq(q
->queue_lock
);
456 spin_lock(&lo
->queue_lock
);
459 spin_unlock(&lo
->queue_lock
);
460 printk(KERN_ERR
"%s: failed between accept and semaphore, file lost\n",
461 lo
->disk
->disk_name
);
463 nbd_end_request(req
);
464 spin_lock_irq(q
->queue_lock
);
468 list_add(&req
->queuelist
, &lo
->queue_head
);
469 spin_unlock(&lo
->queue_lock
);
471 if (nbd_send_req(lo
, req
) != 0) {
472 printk(KERN_ERR
"%s: Request send failed\n",
473 lo
->disk
->disk_name
);
474 if (nbd_find_request(lo
, (char *)&req
) != NULL
) {
475 /* we still own req */
477 nbd_end_request(req
);
478 } else /* we're racing with nbd_clear_que */
479 printk(KERN_DEBUG
"nbd: can't find req\n");
482 spin_lock_irq(q
->queue_lock
);
487 spin_unlock(q
->queue_lock
);
488 nbd_end_request(req
);
489 spin_lock(q
->queue_lock
);
494 static int nbd_ioctl(struct inode
*inode
, struct file
*file
,
495 unsigned int cmd
, unsigned long arg
)
497 struct nbd_device
*lo
= inode
->i_bdev
->bd_disk
->private_data
;
499 struct request sreq
;
501 if (!capable(CAP_SYS_ADMIN
))
504 BUG_ON(lo
->magic
!= LO_MAGIC
);
506 /* Anyone capable of this syscall can do *real bad* things */
507 dprintk(DBG_IOCTL
, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
508 lo
->disk
->disk_name
, ioctl_cmd_to_ascii(cmd
), cmd
, arg
);
512 printk(KERN_INFO
"%s: NBD_DISCONNECT\n", lo
->disk
->disk_name
);
513 sreq
.flags
= REQ_SPECIAL
;
514 nbd_cmd(&sreq
) = NBD_CMD_DISC
;
516 * Set these to sane values in case server implementation
517 * fails to check the request type first and also to keep
518 * debugging output cleaner.
524 nbd_send_req(lo
, &sreq
);
532 spin_lock(&lo
->queue_lock
);
535 spin_unlock(&lo
->queue_lock
);
537 spin_lock(&lo
->queue_lock
);
538 if (!list_empty(&lo
->queue_head
)) {
539 printk(KERN_ERR
"nbd: disconnect: some requests are in progress -> please try again.\n");
542 spin_unlock(&lo
->queue_lock
);
552 inode
= file
->f_dentry
->d_inode
;
553 if (S_ISSOCK(inode
->i_mode
)) {
555 lo
->sock
= SOCKET_I(inode
);
562 case NBD_SET_BLKSIZE
:
564 lo
->bytesize
&= ~(lo
->blksize
-1);
565 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
566 set_blocksize(inode
->i_bdev
, lo
->blksize
);
567 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
570 lo
->bytesize
= arg
& ~(lo
->blksize
-1);
571 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
572 set_blocksize(inode
->i_bdev
, lo
->blksize
);
573 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
575 case NBD_SET_SIZE_BLOCKS
:
576 lo
->bytesize
= ((u64
) arg
) * lo
->blksize
;
577 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
578 set_blocksize(inode
->i_bdev
, lo
->blksize
);
579 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
585 /* on return tidy up in case we have a signal */
586 /* Forcibly shutdown the socket causing all listeners
589 * FIXME: This code is duplicated from sys_shutdown, but
590 * there should be a more generic interface rather than
591 * calling socket ops directly here */
594 printk(KERN_WARNING
"%s: shutting down socket\n",
595 lo
->disk
->disk_name
);
596 lo
->sock
->ops
->shutdown(lo
->sock
,
597 SEND_SHUTDOWN
|RCV_SHUTDOWN
);
601 spin_lock(&lo
->queue_lock
);
604 spin_unlock(&lo
->queue_lock
);
606 printk(KERN_WARNING
"%s: queue cleared\n", lo
->disk
->disk_name
);
609 return lo
->harderror
;
614 return 0; /* probably should be error, but that would
615 * break "nbd-client -d", so just return 0 */
620 case NBD_PRINT_DEBUG
:
621 printk(KERN_INFO
"%s: next = %p, prev = %p, head = %p\n",
622 inode
->i_bdev
->bd_disk
->disk_name
,
623 lo
->queue_head
.next
, lo
->queue_head
.prev
,
630 static struct block_device_operations nbd_fops
=
632 .owner
= THIS_MODULE
,
637 * And here should be modules and kernel interface
638 * (Just smiley confuses emacs :-)
641 static int __init
nbd_init(void)
646 if (sizeof(struct nbd_request
) != 28) {
647 printk(KERN_CRIT
"nbd: sizeof nbd_request needs to be 28 in order to work!\n" );
651 if (nbds_max
> MAX_NBD
) {
652 printk(KERN_CRIT
"nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD
,
657 for (i
= 0; i
< nbds_max
; i
++) {
658 struct gendisk
*disk
= alloc_disk(1);
661 nbd_dev
[i
].disk
= disk
;
663 * The new linux 2.5 block layer implementation requires
664 * every gendisk to have its very own request_queue struct.
665 * These structs are big so we dynamically allocate them.
667 disk
->queue
= blk_init_queue(do_nbd_request
, &nbd_lock
);
674 if (register_blkdev(NBD_MAJOR
, "nbd")) {
679 printk(KERN_INFO
"nbd: registered device at major %d\n", NBD_MAJOR
);
680 dprintk(DBG_INIT
, "nbd: debugflags=0x%x\n", debugflags
);
683 for (i
= 0; i
< nbds_max
; i
++) {
684 struct gendisk
*disk
= nbd_dev
[i
].disk
;
685 nbd_dev
[i
].file
= NULL
;
686 nbd_dev
[i
].magic
= LO_MAGIC
;
687 nbd_dev
[i
].flags
= 0;
688 spin_lock_init(&nbd_dev
[i
].queue_lock
);
689 INIT_LIST_HEAD(&nbd_dev
[i
].queue_head
);
690 init_MUTEX(&nbd_dev
[i
].tx_lock
);
691 nbd_dev
[i
].blksize
= 1024;
692 nbd_dev
[i
].bytesize
= 0x7ffffc00ULL
<< 10; /* 2TB */
693 disk
->major
= NBD_MAJOR
;
694 disk
->first_minor
= i
;
695 disk
->fops
= &nbd_fops
;
696 disk
->private_data
= &nbd_dev
[i
];
697 disk
->flags
|= GENHD_FL_SUPPRESS_PARTITION_INFO
;
698 sprintf(disk
->disk_name
, "nbd%d", i
);
699 sprintf(disk
->devfs_name
, "nbd/%d", i
);
700 set_capacity(disk
, 0x7ffffc00ULL
<< 1); /* 2 TB */
707 blk_cleanup_queue(nbd_dev
[i
].disk
->queue
);
708 put_disk(nbd_dev
[i
].disk
);
713 static void __exit
nbd_cleanup(void)
716 for (i
= 0; i
< nbds_max
; i
++) {
717 struct gendisk
*disk
= nbd_dev
[i
].disk
;
718 nbd_dev
[i
].magic
= 0;
721 blk_cleanup_queue(disk
->queue
);
726 unregister_blkdev(NBD_MAJOR
, "nbd");
727 printk(KERN_INFO
"nbd: unregistered device at major %d\n", NBD_MAJOR
);
730 module_init(nbd_init
);
731 module_exit(nbd_cleanup
);
733 MODULE_DESCRIPTION("Network Block Device");
734 MODULE_LICENSE("GPL");
736 module_param(nbds_max
, int, 0444);
737 MODULE_PARM_DESC(nbds_max
, "How many network block devices to initialize.");
739 module_param(debugflags
, int, 0644);
740 MODULE_PARM_DESC(debugflags
, "flags for controlling debug output");