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 verify_area 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 spinlock_t nbd_lock = SPIN_LOCK_UNLOCKED;
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;
131 struct nbd_device *lo = req->rq_disk->private_data;
134 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
135 req, uptodate? "done": "failed");
137 spin_lock(&lo->queue_lock);
138 while (req->ref_count > 1) { /* still in send */
139 spin_unlock(&lo->queue_lock);
140 printk(KERN_DEBUG "%s: request %p still in use (%d), waiting\n",
141 lo->disk->disk_name, req, req->ref_count);
142 set_current_state(TASK_UNINTERRUPTIBLE);
143 schedule_timeout(HZ); /* wait a second */
144 spin_lock(&lo->queue_lock);
146 spin_unlock(&lo->queue_lock);
148 spin_lock_irqsave(q->queue_lock, flags);
149 if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
150 end_that_request_last(req);
152 spin_unlock_irqrestore(q->queue_lock, flags);
156 * Send or receive packet.
158 static int sock_xmit(struct socket *sock, int send, void *buf, int size,
167 /* Allow interception of SIGKILL only
168 * Don't allow other signals to interrupt the transmission */
169 spin_lock_irqsave(¤t->sighand->siglock, flags);
170 oldset = current->blocked;
171 sigfillset(¤t->blocked);
172 sigdelsetmask(¤t->blocked, sigmask(SIGKILL));
174 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
177 sock->sk->sk_allocation = GFP_NOIO;
182 msg.msg_control = NULL;
183 msg.msg_controllen = 0;
185 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
188 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
190 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
192 if (signal_pending(current)) {
194 spin_lock_irqsave(¤t->sighand->siglock, flags);
195 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
196 current->pid, current->comm,
197 dequeue_signal(current, ¤t->blocked, &info));
198 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
205 result = -EPIPE; /* short read */
212 spin_lock_irqsave(¤t->sighand->siglock, flags);
213 current->blocked = oldset;
215 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
220 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
224 void *kaddr = kmap(bvec->bv_page);
225 result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len,
227 kunmap(bvec->bv_page);
231 void nbd_send_req(struct nbd_device *lo, struct request *req)
233 int result, i, flags;
234 struct nbd_request request;
235 unsigned long size = req->nr_sectors << 9;
236 struct socket *sock = lo->sock;
238 request.magic = htonl(NBD_REQUEST_MAGIC);
239 request.type = htonl(nbd_cmd(req));
240 request.from = cpu_to_be64((u64) req->sector << 9);
241 request.len = htonl(size);
242 memcpy(request.handle, &req, sizeof(req));
246 if (!sock || !lo->sock) {
247 printk(KERN_ERR "%s: Attempted send on closed socket\n",
248 lo->disk->disk_name);
252 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
253 lo->disk->disk_name, req,
254 nbdcmd_to_ascii(nbd_cmd(req)),
255 (unsigned long long)req->sector << 9,
256 req->nr_sectors << 9);
257 result = sock_xmit(sock, 1, &request, sizeof(request),
258 (nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0);
260 printk(KERN_ERR "%s: Send control failed (result %d)\n",
261 lo->disk->disk_name, result);
265 if (nbd_cmd(req) == NBD_CMD_WRITE) {
268 * we are really probing at internals to determine
269 * whether to set MSG_MORE or not...
271 rq_for_each_bio(bio, req) {
272 struct bio_vec *bvec;
273 bio_for_each_segment(bvec, bio, i) {
275 if ((i < (bio->bi_vcnt - 1)) || bio->bi_next)
277 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
278 lo->disk->disk_name, req,
280 result = sock_send_bvec(sock, bvec, flags);
282 printk(KERN_ERR "%s: Send data failed (result %d)\n",
298 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
301 struct list_head *tmp;
302 struct request *xreq;
304 memcpy(&xreq, handle, sizeof(xreq));
306 spin_lock(&lo->queue_lock);
307 list_for_each(tmp, &lo->queue_head) {
308 req = list_entry(tmp, struct request, queuelist);
311 list_del_init(&req->queuelist);
312 spin_unlock(&lo->queue_lock);
315 spin_unlock(&lo->queue_lock);
319 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
322 void *kaddr = kmap(bvec->bv_page);
323 result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len,
325 kunmap(bvec->bv_page);
329 /* NULL returned = something went wrong, inform userspace */
330 struct request *nbd_read_stat(struct nbd_device *lo)
333 struct nbd_reply reply;
335 struct socket *sock = lo->sock;
338 result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL);
340 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
341 lo->disk->disk_name, result);
344 req = nbd_find_request(lo, reply.handle);
346 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
347 lo->disk->disk_name, reply.handle);
352 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
353 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
355 (unsigned long)ntohl(reply.magic));
359 if (ntohl(reply.error)) {
360 printk(KERN_ERR "%s: Other side returned error (%d)\n",
361 lo->disk->disk_name, ntohl(reply.error));
366 dprintk(DBG_RX, "%s: request %p: got reply\n",
367 lo->disk->disk_name, req);
368 if (nbd_cmd(req) == NBD_CMD_READ) {
371 rq_for_each_bio(bio, req) {
372 struct bio_vec *bvec;
373 bio_for_each_segment(bvec, bio, i) {
374 result = sock_recv_bvec(sock, bvec);
376 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
381 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
382 lo->disk->disk_name, req, bvec->bv_len);
388 lo->harderror = result;
392 void nbd_do_it(struct nbd_device *lo)
396 BUG_ON(lo->magic != LO_MAGIC);
398 while ((req = nbd_read_stat(lo)) != NULL)
399 nbd_end_request(req);
403 void nbd_clear_que(struct nbd_device *lo)
407 BUG_ON(lo->magic != LO_MAGIC);
411 spin_lock(&lo->queue_lock);
412 if (!list_empty(&lo->queue_head)) {
413 req = list_entry(lo->queue_head.next, struct request, queuelist);
414 list_del_init(&req->queuelist);
416 spin_unlock(&lo->queue_lock);
419 nbd_end_request(req);
425 * We always wait for result of write, for now. It would be nice to make it optional
427 * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
428 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
431 static void do_nbd_request(request_queue_t * q)
435 while ((req = elv_next_request(q)) != NULL) {
436 struct nbd_device *lo;
438 blkdev_dequeue_request(req);
439 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%lx)\n",
440 req->rq_disk->disk_name, req, req->flags);
442 if (!(req->flags & REQ_CMD))
445 lo = req->rq_disk->private_data;
447 BUG_ON(lo->magic != LO_MAGIC);
450 printk(KERN_ERR "%s: Request when not-ready\n",
451 lo->disk->disk_name);
454 nbd_cmd(req) = NBD_CMD_READ;
455 if (rq_data_dir(req) == WRITE) {
456 nbd_cmd(req) = NBD_CMD_WRITE;
457 if (lo->flags & NBD_READ_ONLY) {
458 printk(KERN_ERR "%s: Write on read-only\n",
459 lo->disk->disk_name);
465 spin_unlock_irq(q->queue_lock);
467 spin_lock(&lo->queue_lock);
470 spin_unlock(&lo->queue_lock);
471 printk(KERN_ERR "%s: failed between accept and semaphore, file lost\n",
472 lo->disk->disk_name);
474 nbd_end_request(req);
475 spin_lock_irq(q->queue_lock);
479 list_add(&req->queuelist, &lo->queue_head);
480 req->ref_count++; /* make sure req does not get freed */
481 spin_unlock(&lo->queue_lock);
483 nbd_send_req(lo, req);
486 printk(KERN_ERR "%s: Request send failed\n",
487 lo->disk->disk_name);
488 spin_lock(&lo->queue_lock);
489 list_del_init(&req->queuelist);
491 spin_unlock(&lo->queue_lock);
492 nbd_end_request(req);
493 spin_lock_irq(q->queue_lock);
497 spin_lock(&lo->queue_lock);
499 spin_unlock(&lo->queue_lock);
500 spin_lock_irq(q->queue_lock);
505 spin_unlock(q->queue_lock);
506 nbd_end_request(req);
507 spin_lock(q->queue_lock);
512 static int nbd_ioctl(struct inode *inode, struct file *file,
513 unsigned int cmd, unsigned long arg)
515 struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
517 struct request sreq ;
519 if (!capable(CAP_SYS_ADMIN))
522 BUG_ON(lo->magic != LO_MAGIC);
524 /* Anyone capable of this syscall can do *real bad* things */
525 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
526 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
530 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
531 sreq.flags = REQ_SPECIAL;
532 nbd_cmd(&sreq) = NBD_CMD_DISC;
534 * Set these to sane values in case server implementation
535 * fails to check the request type first and also to keep
536 * debugging output cleaner.
542 nbd_send_req(lo, &sreq);
550 spin_lock(&lo->queue_lock);
553 spin_unlock(&lo->queue_lock);
555 spin_lock(&lo->queue_lock);
556 if (!list_empty(&lo->queue_head)) {
557 printk(KERN_ERR "nbd: disconnect: some requests are in progress -> please try again.\n");
560 spin_unlock(&lo->queue_lock);
570 inode = file->f_dentry->d_inode;
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_SIZE_BLOCKS:
594 lo->bytesize = ((u64) arg) * lo->blksize;
595 inode->i_bdev->bd_inode->i_size = lo->bytesize;
596 set_blocksize(inode->i_bdev, lo->blksize);
597 set_capacity(lo->disk, lo->bytesize >> 9);
603 /* on return tidy up in case we have a signal */
604 /* Forcibly shutdown the socket causing all listeners
607 * FIXME: This code is duplicated from sys_shutdown, but
608 * there should be a more generic interface rather than
609 * calling socket ops directly here */
612 printk(KERN_WARNING "%s: shutting down socket\n",
613 lo->disk->disk_name);
614 lo->sock->ops->shutdown(lo->sock,
615 SEND_SHUTDOWN|RCV_SHUTDOWN);
619 spin_lock(&lo->queue_lock);
622 spin_unlock(&lo->queue_lock);
624 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
627 return lo->harderror;
632 return 0; /* probably should be error, but that would
633 * break "nbd-client -d", so just return 0 */
638 case NBD_PRINT_DEBUG:
639 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
640 inode->i_bdev->bd_disk->disk_name,
641 lo->queue_head.next, lo->queue_head.prev,
648 static struct block_device_operations nbd_fops =
650 .owner = THIS_MODULE,
655 * And here should be modules and kernel interface
656 * (Just smiley confuses emacs :-)
659 static int __init nbd_init(void)
664 if (sizeof(struct nbd_request) != 28) {
665 printk(KERN_CRIT "nbd: sizeof nbd_request needs to be 28 in order to work!\n" );
669 for (i = 0; i < MAX_NBD; i++) {
670 struct gendisk *disk = alloc_disk(1);
673 nbd_dev[i].disk = disk;
675 * The new linux 2.5 block layer implementation requires
676 * every gendisk to have its very own request_queue struct.
677 * These structs are big so we dynamically allocate them.
679 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
686 if (register_blkdev(NBD_MAJOR, "nbd")) {
691 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
692 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
695 for (i = 0; i < MAX_NBD; i++) {
696 struct gendisk *disk = nbd_dev[i].disk;
697 nbd_dev[i].file = NULL;
698 nbd_dev[i].magic = LO_MAGIC;
699 nbd_dev[i].flags = 0;
700 spin_lock_init(&nbd_dev[i].queue_lock);
701 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
702 init_MUTEX(&nbd_dev[i].tx_lock);
703 nbd_dev[i].blksize = 1024;
704 nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */
705 disk->major = NBD_MAJOR;
706 disk->first_minor = i;
707 disk->fops = &nbd_fops;
708 disk->private_data = &nbd_dev[i];
709 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
710 sprintf(disk->disk_name, "nbd%d", i);
711 sprintf(disk->devfs_name, "nbd/%d", i);
712 set_capacity(disk, 0x7ffffc00ULL << 1); /* 2 TB */
719 blk_cleanup_queue(nbd_dev[i].disk->queue);
720 put_disk(nbd_dev[i].disk);
725 static void __exit nbd_cleanup(void)
728 for (i = 0; i < MAX_NBD; i++) {
729 struct gendisk *disk = nbd_dev[i].disk;
732 blk_cleanup_queue(disk->queue);
737 unregister_blkdev(NBD_MAJOR, "nbd");
738 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
741 module_init(nbd_init);
742 module_exit(nbd_cleanup);
744 MODULE_DESCRIPTION("Network Block Device");
745 MODULE_LICENSE("GPL");
748 MODULE_PARM(debugflags, "i");
749 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");