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 struct request_queue *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,
125 sigset_t blocked, oldset;
127 /* Allow interception of SIGKILL only
128 * Don't allow other signals to interrupt the transmission */
129 siginitsetinv(&blocked, sigmask(SIGKILL));
130 sigprocmask(SIG_SETMASK, &blocked, &oldset);
133 sock->sk->sk_allocation = GFP_NOIO;
138 msg.msg_control = NULL;
139 msg.msg_controllen = 0;
140 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
143 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
145 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
147 if (signal_pending(current)) {
149 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
150 current->pid, current->comm,
151 dequeue_signal_lock(current, ¤t->blocked, &info));
158 result = -EPIPE; /* short read */
165 sigprocmask(SIG_SETMASK, &oldset, NULL);
170 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
174 void *kaddr = kmap(bvec->bv_page);
175 result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len,
177 kunmap(bvec->bv_page);
181 static int nbd_send_req(struct nbd_device *lo, struct request *req)
184 struct nbd_request request;
185 unsigned long size = req->nr_sectors << 9;
186 struct socket *sock = lo->sock;
188 request.magic = htonl(NBD_REQUEST_MAGIC);
189 request.type = htonl(nbd_cmd(req));
190 request.from = cpu_to_be64((u64) req->sector << 9);
191 request.len = htonl(size);
192 memcpy(request.handle, &req, sizeof(req));
194 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
195 lo->disk->disk_name, req,
196 nbdcmd_to_ascii(nbd_cmd(req)),
197 (unsigned long long)req->sector << 9,
198 req->nr_sectors << 9);
199 result = sock_xmit(sock, 1, &request, sizeof(request),
200 (nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0);
202 printk(KERN_ERR "%s: Send control failed (result %d)\n",
203 lo->disk->disk_name, result);
207 if (nbd_cmd(req) == NBD_CMD_WRITE) {
208 struct req_iterator iter;
209 struct bio_vec *bvec;
211 * we are really probing at internals to determine
212 * whether to set MSG_MORE or not...
214 rq_for_each_segment(bvec, req, iter) {
216 if (!rq_iter_last(req, iter))
218 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
219 lo->disk->disk_name, req, bvec->bv_len);
220 result = sock_send_bvec(sock, bvec, flags);
222 printk(KERN_ERR "%s: Send data failed (result %d)\n",
223 lo->disk->disk_name, result);
234 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
237 struct list_head *tmp;
238 struct request *xreq;
241 memcpy(&xreq, handle, sizeof(xreq));
243 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
247 spin_lock(&lo->queue_lock);
248 list_for_each(tmp, &lo->queue_head) {
249 req = list_entry(tmp, struct request, queuelist);
252 list_del_init(&req->queuelist);
253 spin_unlock(&lo->queue_lock);
256 spin_unlock(&lo->queue_lock);
264 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
267 void *kaddr = kmap(bvec->bv_page);
268 result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len,
270 kunmap(bvec->bv_page);
274 /* NULL returned = something went wrong, inform userspace */
275 static struct request *nbd_read_stat(struct nbd_device *lo)
278 struct nbd_reply reply;
280 struct socket *sock = lo->sock;
283 result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL);
285 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
286 lo->disk->disk_name, result);
290 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
291 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
293 (unsigned long)ntohl(reply.magic));
298 req = nbd_find_request(lo, reply.handle);
299 if (unlikely(IS_ERR(req))) {
300 result = PTR_ERR(req);
301 if (result != -ENOENT)
304 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
305 lo->disk->disk_name, reply.handle);
310 if (ntohl(reply.error)) {
311 printk(KERN_ERR "%s: Other side returned error (%d)\n",
312 lo->disk->disk_name, ntohl(reply.error));
317 dprintk(DBG_RX, "%s: request %p: got reply\n",
318 lo->disk->disk_name, req);
319 if (nbd_cmd(req) == NBD_CMD_READ) {
320 struct req_iterator iter;
321 struct bio_vec *bvec;
323 rq_for_each_segment(bvec, req, iter) {
324 result = sock_recv_bvec(sock, bvec);
326 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
327 lo->disk->disk_name, result);
331 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
332 lo->disk->disk_name, req, bvec->bv_len);
337 lo->harderror = result;
341 static ssize_t pid_show(struct gendisk *disk, char *page)
343 return sprintf(page, "%ld\n",
344 (long) ((struct nbd_device *)disk->private_data)->pid);
347 static struct disk_attribute pid_attr = {
348 .attr = { .name = "pid", .mode = S_IRUGO },
352 static int nbd_do_it(struct nbd_device *lo)
357 BUG_ON(lo->magic != LO_MAGIC);
359 lo->pid = current->pid;
360 ret = sysfs_create_file(&lo->disk->kobj, &pid_attr.attr);
362 printk(KERN_ERR "nbd: sysfs_create_file failed!");
366 while ((req = nbd_read_stat(lo)) != NULL)
367 nbd_end_request(req);
369 sysfs_remove_file(&lo->disk->kobj, &pid_attr.attr);
373 static void nbd_clear_que(struct nbd_device *lo)
377 BUG_ON(lo->magic != LO_MAGIC);
380 * Because we have set lo->sock to NULL under the tx_lock, all
381 * modifications to the list must have completed by now. For
382 * the same reason, the active_req must be NULL.
384 * As a consequence, we don't need to take the spin lock while
385 * purging the list here.
388 BUG_ON(lo->active_req);
390 while (!list_empty(&lo->queue_head)) {
391 req = list_entry(lo->queue_head.next, struct request,
393 list_del_init(&req->queuelist);
395 nbd_end_request(req);
400 * We always wait for result of write, for now. It would be nice to make it optional
402 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
403 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
406 static void do_nbd_request(struct request_queue * q)
410 while ((req = elv_next_request(q)) != NULL) {
411 struct nbd_device *lo;
413 blkdev_dequeue_request(req);
414 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
415 req->rq_disk->disk_name, req, req->cmd_type);
417 if (!blk_fs_request(req))
420 lo = req->rq_disk->private_data;
422 BUG_ON(lo->magic != LO_MAGIC);
424 nbd_cmd(req) = NBD_CMD_READ;
425 if (rq_data_dir(req) == WRITE) {
426 nbd_cmd(req) = NBD_CMD_WRITE;
427 if (lo->flags & NBD_READ_ONLY) {
428 printk(KERN_ERR "%s: Write on read-only\n",
429 lo->disk->disk_name);
435 spin_unlock_irq(q->queue_lock);
437 mutex_lock(&lo->tx_lock);
438 if (unlikely(!lo->sock)) {
439 mutex_unlock(&lo->tx_lock);
440 printk(KERN_ERR "%s: Attempted send on closed socket\n",
441 lo->disk->disk_name);
443 nbd_end_request(req);
444 spin_lock_irq(q->queue_lock);
448 lo->active_req = req;
450 if (nbd_send_req(lo, req) != 0) {
451 printk(KERN_ERR "%s: Request send failed\n",
452 lo->disk->disk_name);
454 nbd_end_request(req);
456 spin_lock(&lo->queue_lock);
457 list_add(&req->queuelist, &lo->queue_head);
458 spin_unlock(&lo->queue_lock);
461 lo->active_req = NULL;
462 mutex_unlock(&lo->tx_lock);
463 wake_up_all(&lo->active_wq);
465 spin_lock_irq(q->queue_lock);
470 spin_unlock(q->queue_lock);
471 nbd_end_request(req);
472 spin_lock(q->queue_lock);
477 static int nbd_ioctl(struct inode *inode, struct file *file,
478 unsigned int cmd, unsigned long arg)
480 struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
482 struct request sreq ;
484 if (!capable(CAP_SYS_ADMIN))
487 BUG_ON(lo->magic != LO_MAGIC);
489 /* Anyone capable of this syscall can do *real bad* things */
490 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
491 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
495 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
496 sreq.cmd_type = REQ_TYPE_SPECIAL;
497 nbd_cmd(&sreq) = NBD_CMD_DISC;
499 * Set these to sane values in case server implementation
500 * fails to check the request type first and also to keep
501 * debugging output cleaner.
507 nbd_send_req(lo, &sreq);
512 mutex_lock(&lo->tx_lock);
514 mutex_unlock(&lo->tx_lock);
518 BUG_ON(!list_empty(&lo->queue_head));
528 inode = file->f_path.dentry->d_inode;
529 if (S_ISSOCK(inode->i_mode)) {
531 lo->sock = SOCKET_I(inode);
538 case NBD_SET_BLKSIZE:
540 lo->bytesize &= ~(lo->blksize-1);
541 inode->i_bdev->bd_inode->i_size = lo->bytesize;
542 set_blocksize(inode->i_bdev, lo->blksize);
543 set_capacity(lo->disk, lo->bytesize >> 9);
546 lo->bytesize = arg & ~(lo->blksize-1);
547 inode->i_bdev->bd_inode->i_size = lo->bytesize;
548 set_blocksize(inode->i_bdev, lo->blksize);
549 set_capacity(lo->disk, lo->bytesize >> 9);
551 case NBD_SET_SIZE_BLOCKS:
552 lo->bytesize = ((u64) arg) * lo->blksize;
553 inode->i_bdev->bd_inode->i_size = lo->bytesize;
554 set_blocksize(inode->i_bdev, lo->blksize);
555 set_capacity(lo->disk, lo->bytesize >> 9);
560 error = nbd_do_it(lo);
563 /* on return tidy up in case we have a signal */
564 /* Forcibly shutdown the socket causing all listeners
567 * FIXME: This code is duplicated from sys_shutdown, but
568 * there should be a more generic interface rather than
569 * calling socket ops directly here */
570 mutex_lock(&lo->tx_lock);
572 printk(KERN_WARNING "%s: shutting down socket\n",
573 lo->disk->disk_name);
574 lo->sock->ops->shutdown(lo->sock,
575 SEND_SHUTDOWN|RCV_SHUTDOWN);
578 mutex_unlock(&lo->tx_lock);
582 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
585 return lo->harderror;
588 * This is for compatibility only. The queue is always cleared
589 * by NBD_DO_IT or NBD_CLEAR_SOCK.
591 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
593 case NBD_PRINT_DEBUG:
594 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
595 inode->i_bdev->bd_disk->disk_name,
596 lo->queue_head.next, lo->queue_head.prev,
603 static struct block_device_operations nbd_fops =
605 .owner = THIS_MODULE,
610 * And here should be modules and kernel interface
611 * (Just smiley confuses emacs :-)
614 static int __init nbd_init(void)
619 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
621 if (nbds_max > MAX_NBD) {
622 printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD,
627 for (i = 0; i < nbds_max; i++) {
628 struct gendisk *disk = alloc_disk(1);
631 nbd_dev[i].disk = disk;
633 * The new linux 2.5 block layer implementation requires
634 * every gendisk to have its very own request_queue struct.
635 * These structs are big so we dynamically allocate them.
637 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
644 if (register_blkdev(NBD_MAJOR, "nbd")) {
649 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
650 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
652 for (i = 0; i < nbds_max; i++) {
653 struct gendisk *disk = nbd_dev[i].disk;
654 nbd_dev[i].file = NULL;
655 nbd_dev[i].magic = LO_MAGIC;
656 nbd_dev[i].flags = 0;
657 spin_lock_init(&nbd_dev[i].queue_lock);
658 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
659 mutex_init(&nbd_dev[i].tx_lock);
660 init_waitqueue_head(&nbd_dev[i].active_wq);
661 nbd_dev[i].blksize = 1024;
662 nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */
663 disk->major = NBD_MAJOR;
664 disk->first_minor = i;
665 disk->fops = &nbd_fops;
666 disk->private_data = &nbd_dev[i];
667 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
668 sprintf(disk->disk_name, "nbd%d", i);
669 set_capacity(disk, 0x7ffffc00ULL << 1); /* 2 TB */
676 blk_cleanup_queue(nbd_dev[i].disk->queue);
677 put_disk(nbd_dev[i].disk);
682 static void __exit nbd_cleanup(void)
685 for (i = 0; i < nbds_max; i++) {
686 struct gendisk *disk = nbd_dev[i].disk;
687 nbd_dev[i].magic = 0;
690 blk_cleanup_queue(disk->queue);
694 unregister_blkdev(NBD_MAJOR, "nbd");
695 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
698 module_init(nbd_init);
699 module_exit(nbd_cleanup);
701 MODULE_DESCRIPTION("Network Block Device");
702 MODULE_LICENSE("GPL");
704 module_param(nbds_max, int, 0444);
705 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
707 module_param(debugflags, int, 0644);
708 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");