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>
31 #include <linux/net.h>
33 #include <asm/uaccess.h>
34 #include <asm/system.h>
35 #include <asm/types.h>
37 #include <linux/nbd.h>
39 #define LO_MAGIC 0x68797548
42 #define dprintk(flags, fmt...)
44 #define dprintk(flags, fmt...) do { \
45 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
47 #define DBG_IOCTL 0x0004
48 #define DBG_INIT 0x0010
49 #define DBG_EXIT 0x0020
50 #define DBG_BLKDEV 0x0100
53 static unsigned int debugflags
;
56 static unsigned int nbds_max
= 16;
57 static struct nbd_device nbd_dev
[MAX_NBD
];
60 * Use just one lock (or at most 1 per NIC). Two arguments for this:
61 * 1. Each NIC is essentially a synchronization point for all servers
62 * accessed through that NIC so there's no need to have more locks
64 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
65 * down each lock to the point where they're actually slower than just
67 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
69 static DEFINE_SPINLOCK(nbd_lock
);
72 static const char *ioctl_cmd_to_ascii(int cmd
)
75 case NBD_SET_SOCK
: return "set-sock";
76 case NBD_SET_BLKSIZE
: return "set-blksize";
77 case NBD_SET_SIZE
: return "set-size";
78 case NBD_DO_IT
: return "do-it";
79 case NBD_CLEAR_SOCK
: return "clear-sock";
80 case NBD_CLEAR_QUE
: return "clear-que";
81 case NBD_PRINT_DEBUG
: return "print-debug";
82 case NBD_SET_SIZE_BLOCKS
: return "set-size-blocks";
83 case NBD_DISCONNECT
: return "disconnect";
84 case BLKROSET
: return "set-read-only";
85 case BLKFLSBUF
: return "flush-buffer-cache";
90 static const char *nbdcmd_to_ascii(int cmd
)
93 case NBD_CMD_READ
: return "read";
94 case NBD_CMD_WRITE
: return "write";
95 case NBD_CMD_DISC
: return "disconnect";
101 static void nbd_end_request(struct request
*req
)
103 int uptodate
= (req
->errors
== 0) ? 1 : 0;
104 struct request_queue
*q
= req
->q
;
107 dprintk(DBG_BLKDEV
, "%s: request %p: %s\n", req
->rq_disk
->disk_name
,
108 req
, uptodate
? "done": "failed");
110 spin_lock_irqsave(q
->queue_lock
, flags
);
111 if (!end_that_request_first(req
, uptodate
, req
->nr_sectors
)) {
112 end_that_request_last(req
, uptodate
);
114 spin_unlock_irqrestore(q
->queue_lock
, flags
);
117 static void sock_shutdown(struct nbd_device
*lo
, int lock
)
119 /* Forcibly shutdown the socket causing all listeners
122 * FIXME: This code is duplicated from sys_shutdown, but
123 * there should be a more generic interface rather than
124 * calling socket ops directly here */
126 mutex_lock(&lo
->tx_lock
);
128 printk(KERN_WARNING
"%s: shutting down socket\n",
129 lo
->disk
->disk_name
);
130 kernel_sock_shutdown(lo
->sock
, SHUT_RDWR
);
134 mutex_unlock(&lo
->tx_lock
);
137 static void nbd_xmit_timeout(unsigned long arg
)
139 struct task_struct
*task
= (struct task_struct
*)arg
;
141 printk(KERN_WARNING
"nbd: killing hung xmit (%s, pid: %d)\n",
142 task
->comm
, task
->pid
);
143 force_sig(SIGKILL
, task
);
147 * Send or receive packet.
149 static int sock_xmit(struct nbd_device
*lo
, int send
, void *buf
, int size
,
152 struct socket
*sock
= lo
->sock
;
156 sigset_t blocked
, oldset
;
158 /* Allow interception of SIGKILL only
159 * Don't allow other signals to interrupt the transmission */
160 siginitsetinv(&blocked
, sigmask(SIGKILL
));
161 sigprocmask(SIG_SETMASK
, &blocked
, &oldset
);
164 sock
->sk
->sk_allocation
= GFP_NOIO
;
169 msg
.msg_control
= NULL
;
170 msg
.msg_controllen
= 0;
171 msg
.msg_flags
= msg_flags
| MSG_NOSIGNAL
;
174 struct timer_list ti
;
176 if (lo
->xmit_timeout
) {
178 ti
.function
= nbd_xmit_timeout
;
179 ti
.data
= (unsigned long)current
;
180 ti
.expires
= jiffies
+ lo
->xmit_timeout
;
183 result
= kernel_sendmsg(sock
, &msg
, &iov
, 1, size
);
184 if (lo
->xmit_timeout
)
187 result
= kernel_recvmsg(sock
, &msg
, &iov
, 1, size
, 0);
189 if (signal_pending(current
)) {
191 printk(KERN_WARNING
"nbd (pid %d: %s) got signal %d\n",
192 task_pid_nr(current
), current
->comm
,
193 dequeue_signal_lock(current
, ¤t
->blocked
, &info
));
195 sock_shutdown(lo
, !send
);
201 result
= -EPIPE
; /* short read */
208 sigprocmask(SIG_SETMASK
, &oldset
, NULL
);
213 static inline int sock_send_bvec(struct nbd_device
*lo
, struct bio_vec
*bvec
,
217 void *kaddr
= kmap(bvec
->bv_page
);
218 result
= sock_xmit(lo
, 1, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
, flags
);
219 kunmap(bvec
->bv_page
);
223 /* always call with the tx_lock held */
224 static int nbd_send_req(struct nbd_device
*lo
, struct request
*req
)
227 struct nbd_request request
;
228 unsigned long size
= req
->nr_sectors
<< 9;
230 request
.magic
= htonl(NBD_REQUEST_MAGIC
);
231 request
.type
= htonl(nbd_cmd(req
));
232 request
.from
= cpu_to_be64((u64
) req
->sector
<< 9);
233 request
.len
= htonl(size
);
234 memcpy(request
.handle
, &req
, sizeof(req
));
236 dprintk(DBG_TX
, "%s: request %p: sending control (%s@%llu,%luB)\n",
237 lo
->disk
->disk_name
, req
,
238 nbdcmd_to_ascii(nbd_cmd(req
)),
239 (unsigned long long)req
->sector
<< 9,
240 req
->nr_sectors
<< 9);
241 result
= sock_xmit(lo
, 1, &request
, sizeof(request
),
242 (nbd_cmd(req
) == NBD_CMD_WRITE
) ? MSG_MORE
: 0);
244 printk(KERN_ERR
"%s: Send control failed (result %d)\n",
245 lo
->disk
->disk_name
, result
);
249 if (nbd_cmd(req
) == NBD_CMD_WRITE
) {
250 struct req_iterator iter
;
251 struct bio_vec
*bvec
;
253 * we are really probing at internals to determine
254 * whether to set MSG_MORE or not...
256 rq_for_each_segment(bvec
, req
, iter
) {
258 if (!rq_iter_last(req
, iter
))
260 dprintk(DBG_TX
, "%s: request %p: sending %d bytes data\n",
261 lo
->disk
->disk_name
, req
, bvec
->bv_len
);
262 result
= sock_send_bvec(lo
, bvec
, flags
);
264 printk(KERN_ERR
"%s: Send data failed (result %d)\n",
265 lo
->disk
->disk_name
, result
);
276 static struct request
*nbd_find_request(struct nbd_device
*lo
,
277 struct request
*xreq
)
279 struct request
*req
, *tmp
;
282 err
= wait_event_interruptible(lo
->active_wq
, lo
->active_req
!= xreq
);
286 spin_lock(&lo
->queue_lock
);
287 list_for_each_entry_safe(req
, tmp
, &lo
->queue_head
, queuelist
) {
290 list_del_init(&req
->queuelist
);
291 spin_unlock(&lo
->queue_lock
);
294 spin_unlock(&lo
->queue_lock
);
302 static inline int sock_recv_bvec(struct nbd_device
*lo
, struct bio_vec
*bvec
)
305 void *kaddr
= kmap(bvec
->bv_page
);
306 result
= sock_xmit(lo
, 0, kaddr
+ bvec
->bv_offset
, bvec
->bv_len
,
308 kunmap(bvec
->bv_page
);
312 /* NULL returned = something went wrong, inform userspace */
313 static struct request
*nbd_read_stat(struct nbd_device
*lo
)
316 struct nbd_reply reply
;
320 result
= sock_xmit(lo
, 0, &reply
, sizeof(reply
), MSG_WAITALL
);
322 printk(KERN_ERR
"%s: Receive control failed (result %d)\n",
323 lo
->disk
->disk_name
, result
);
327 if (ntohl(reply
.magic
) != NBD_REPLY_MAGIC
) {
328 printk(KERN_ERR
"%s: Wrong magic (0x%lx)\n",
330 (unsigned long)ntohl(reply
.magic
));
335 req
= nbd_find_request(lo
, *(struct request
**)reply
.handle
);
336 if (unlikely(IS_ERR(req
))) {
337 result
= PTR_ERR(req
);
338 if (result
!= -ENOENT
)
341 printk(KERN_ERR
"%s: Unexpected reply (%p)\n",
342 lo
->disk
->disk_name
, reply
.handle
);
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
) {
357 struct req_iterator iter
;
358 struct bio_vec
*bvec
;
360 rq_for_each_segment(bvec
, req
, iter
) {
361 result
= sock_recv_bvec(lo
, bvec
);
363 printk(KERN_ERR
"%s: Receive data failed (result %d)\n",
364 lo
->disk
->disk_name
, result
);
368 dprintk(DBG_RX
, "%s: request %p: got %d bytes data\n",
369 lo
->disk
->disk_name
, req
, bvec
->bv_len
);
374 lo
->harderror
= result
;
378 static ssize_t
pid_show(struct device
*dev
,
379 struct device_attribute
*attr
, char *buf
)
381 struct gendisk
*disk
= dev_to_disk(dev
);
383 return sprintf(buf
, "%ld\n",
384 (long) ((struct nbd_device
*)disk
->private_data
)->pid
);
387 static struct device_attribute pid_attr
= {
388 .attr
= { .name
= "pid", .mode
= S_IRUGO
, .owner
= THIS_MODULE
},
392 static int nbd_do_it(struct nbd_device
*lo
)
397 BUG_ON(lo
->magic
!= LO_MAGIC
);
399 lo
->pid
= current
->pid
;
400 ret
= sysfs_create_file(&lo
->disk
->dev
.kobj
, &pid_attr
.attr
);
402 printk(KERN_ERR
"nbd: sysfs_create_file failed!");
406 while ((req
= nbd_read_stat(lo
)) != NULL
)
407 nbd_end_request(req
);
409 sysfs_remove_file(&lo
->disk
->dev
.kobj
, &pid_attr
.attr
);
413 static void nbd_clear_que(struct nbd_device
*lo
)
417 BUG_ON(lo
->magic
!= LO_MAGIC
);
420 * Because we have set lo->sock to NULL under the tx_lock, all
421 * modifications to the list must have completed by now. For
422 * the same reason, the active_req must be NULL.
424 * As a consequence, we don't need to take the spin lock while
425 * purging the list here.
428 BUG_ON(lo
->active_req
);
430 while (!list_empty(&lo
->queue_head
)) {
431 req
= list_entry(lo
->queue_head
.next
, struct request
,
433 list_del_init(&req
->queuelist
);
435 nbd_end_request(req
);
441 * We always wait for result of write, for now. It would be nice to make it optional
443 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
444 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
447 static void do_nbd_request(struct request_queue
* q
)
451 while ((req
= elv_next_request(q
)) != NULL
) {
452 struct nbd_device
*lo
;
454 blkdev_dequeue_request(req
);
455 dprintk(DBG_BLKDEV
, "%s: request %p: dequeued (flags=%x)\n",
456 req
->rq_disk
->disk_name
, req
, req
->cmd_type
);
458 if (!blk_fs_request(req
))
461 lo
= req
->rq_disk
->private_data
;
463 BUG_ON(lo
->magic
!= LO_MAGIC
);
465 nbd_cmd(req
) = NBD_CMD_READ
;
466 if (rq_data_dir(req
) == WRITE
) {
467 nbd_cmd(req
) = NBD_CMD_WRITE
;
468 if (lo
->flags
& NBD_READ_ONLY
) {
469 printk(KERN_ERR
"%s: Write on read-only\n",
470 lo
->disk
->disk_name
);
476 spin_unlock_irq(q
->queue_lock
);
478 mutex_lock(&lo
->tx_lock
);
479 if (unlikely(!lo
->sock
)) {
480 mutex_unlock(&lo
->tx_lock
);
481 printk(KERN_ERR
"%s: Attempted send on closed socket\n",
482 lo
->disk
->disk_name
);
484 nbd_end_request(req
);
485 spin_lock_irq(q
->queue_lock
);
489 lo
->active_req
= req
;
491 if (nbd_send_req(lo
, req
) != 0) {
492 printk(KERN_ERR
"%s: Request send failed\n",
493 lo
->disk
->disk_name
);
495 nbd_end_request(req
);
497 spin_lock(&lo
->queue_lock
);
498 list_add(&req
->queuelist
, &lo
->queue_head
);
499 spin_unlock(&lo
->queue_lock
);
502 lo
->active_req
= NULL
;
503 mutex_unlock(&lo
->tx_lock
);
504 wake_up_all(&lo
->active_wq
);
506 spin_lock_irq(q
->queue_lock
);
511 spin_unlock(q
->queue_lock
);
512 nbd_end_request(req
);
513 spin_lock(q
->queue_lock
);
517 static int nbd_ioctl(struct inode
*inode
, struct file
*file
,
518 unsigned int cmd
, unsigned long arg
)
520 struct nbd_device
*lo
= inode
->i_bdev
->bd_disk
->private_data
;
522 struct request sreq
;
524 if (!capable(CAP_SYS_ADMIN
))
527 BUG_ON(lo
->magic
!= LO_MAGIC
);
529 /* Anyone capable of this syscall can do *real bad* things */
530 dprintk(DBG_IOCTL
, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
531 lo
->disk
->disk_name
, ioctl_cmd_to_ascii(cmd
), cmd
, arg
);
535 printk(KERN_INFO
"%s: NBD_DISCONNECT\n", lo
->disk
->disk_name
);
536 sreq
.cmd_type
= REQ_TYPE_SPECIAL
;
537 nbd_cmd(&sreq
) = NBD_CMD_DISC
;
539 * Set these to sane values in case server implementation
540 * fails to check the request type first and also to keep
541 * debugging output cleaner.
547 mutex_lock(&lo
->tx_lock
);
548 nbd_send_req(lo
, &sreq
);
549 mutex_unlock(&lo
->tx_lock
);
554 mutex_lock(&lo
->tx_lock
);
556 mutex_unlock(&lo
->tx_lock
);
560 BUG_ON(!list_empty(&lo
->queue_head
));
570 inode
= file
->f_path
.dentry
->d_inode
;
571 if (S_ISSOCK(inode
->i_mode
)) {
573 lo
->sock
= SOCKET_I(inode
);
580 case NBD_SET_BLKSIZE
:
582 lo
->bytesize
&= ~(lo
->blksize
-1);
583 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
584 set_blocksize(inode
->i_bdev
, lo
->blksize
);
585 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
588 lo
->bytesize
= arg
& ~(lo
->blksize
-1);
589 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
590 set_blocksize(inode
->i_bdev
, lo
->blksize
);
591 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
593 case NBD_SET_TIMEOUT
:
594 lo
->xmit_timeout
= arg
* HZ
;
596 case NBD_SET_SIZE_BLOCKS
:
597 lo
->bytesize
= ((u64
) arg
) * lo
->blksize
;
598 inode
->i_bdev
->bd_inode
->i_size
= lo
->bytesize
;
599 set_blocksize(inode
->i_bdev
, lo
->blksize
);
600 set_capacity(lo
->disk
, lo
->bytesize
>> 9);
605 error
= nbd_do_it(lo
);
608 sock_shutdown(lo
, 1);
612 printk(KERN_WARNING
"%s: queue cleared\n", lo
->disk
->disk_name
);
616 inode
->i_bdev
->bd_inode
->i_size
= 0;
617 set_capacity(lo
->disk
, 0);
618 return lo
->harderror
;
621 * This is for compatibility only. The queue is always cleared
622 * by NBD_DO_IT or NBD_CLEAR_SOCK.
624 BUG_ON(!lo
->sock
&& !list_empty(&lo
->queue_head
));
626 case NBD_PRINT_DEBUG
:
627 printk(KERN_INFO
"%s: next = %p, prev = %p, head = %p\n",
628 inode
->i_bdev
->bd_disk
->disk_name
,
629 lo
->queue_head
.next
, lo
->queue_head
.prev
,
636 static struct block_device_operations nbd_fops
=
638 .owner
= THIS_MODULE
,
643 * And here should be modules and kernel interface
644 * (Just smiley confuses emacs :-)
647 static int __init
nbd_init(void)
652 BUILD_BUG_ON(sizeof(struct nbd_request
) != 28);
654 if (nbds_max
> MAX_NBD
) {
655 printk(KERN_CRIT
"nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD
,
660 for (i
= 0; i
< nbds_max
; i
++) {
661 struct gendisk
*disk
= alloc_disk(1);
664 nbd_dev
[i
].disk
= disk
;
666 * The new linux 2.5 block layer implementation requires
667 * every gendisk to have its very own request_queue struct.
668 * These structs are big so we dynamically allocate them.
670 disk
->queue
= blk_init_queue(do_nbd_request
, &nbd_lock
);
677 if (register_blkdev(NBD_MAJOR
, "nbd")) {
682 printk(KERN_INFO
"nbd: registered device at major %d\n", NBD_MAJOR
);
683 dprintk(DBG_INIT
, "nbd: debugflags=0x%x\n", debugflags
);
685 for (i
= 0; i
< nbds_max
; i
++) {
686 struct gendisk
*disk
= nbd_dev
[i
].disk
;
687 nbd_dev
[i
].file
= NULL
;
688 nbd_dev
[i
].magic
= LO_MAGIC
;
689 nbd_dev
[i
].flags
= 0;
690 spin_lock_init(&nbd_dev
[i
].queue_lock
);
691 INIT_LIST_HEAD(&nbd_dev
[i
].queue_head
);
692 mutex_init(&nbd_dev
[i
].tx_lock
);
693 init_waitqueue_head(&nbd_dev
[i
].active_wq
);
694 nbd_dev
[i
].blksize
= 1024;
695 nbd_dev
[i
].bytesize
= 0;
696 disk
->major
= NBD_MAJOR
;
697 disk
->first_minor
= i
;
698 disk
->fops
= &nbd_fops
;
699 disk
->private_data
= &nbd_dev
[i
];
700 disk
->flags
|= GENHD_FL_SUPPRESS_PARTITION_INFO
;
701 sprintf(disk
->disk_name
, "nbd%d", i
);
702 set_capacity(disk
, 0);
709 blk_cleanup_queue(nbd_dev
[i
].disk
->queue
);
710 put_disk(nbd_dev
[i
].disk
);
715 static void __exit
nbd_cleanup(void)
718 for (i
= 0; i
< nbds_max
; i
++) {
719 struct gendisk
*disk
= nbd_dev
[i
].disk
;
720 nbd_dev
[i
].magic
= 0;
723 blk_cleanup_queue(disk
->queue
);
727 unregister_blkdev(NBD_MAJOR
, "nbd");
728 printk(KERN_INFO
"nbd: unregistered device at major %d\n", NBD_MAJOR
);
731 module_init(nbd_init
);
732 module_exit(nbd_cleanup
);
734 MODULE_DESCRIPTION("Network Block Device");
735 MODULE_LICENSE("GPL");
737 module_param(nbds_max
, int, 0444);
738 MODULE_PARM_DESC(nbds_max
, "How many network block devices to initialize.");
740 module_param(debugflags
, int, 0644);
741 MODULE_PARM_DESC(debugflags
, "flags for controlling debug output");