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, 2008 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/mutex.h>
28 #include <linux/compiler.h>
29 #include <linux/err.h>
30 #include <linux/kernel.h>
31 #include <linux/slab.h>
33 #include <linux/net.h>
34 #include <linux/kthread.h>
36 #include <asm/uaccess.h>
37 #include <asm/system.h>
38 #include <asm/types.h>
40 #include <linux/nbd.h>
42 #define LO_MAGIC 0x68797548
45 #define dprintk(flags, fmt...)
47 #define dprintk(flags, fmt...) do { \
48 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
50 #define DBG_IOCTL 0x0004
51 #define DBG_INIT 0x0010
52 #define DBG_EXIT 0x0020
53 #define DBG_BLKDEV 0x0100
56 static unsigned int debugflags;
59 static unsigned int nbds_max = 16;
60 static struct nbd_device *nbd_dev;
64 * Use just one lock (or at most 1 per NIC). Two arguments for this:
65 * 1. Each NIC is essentially a synchronization point for all servers
66 * accessed through that NIC so there's no need to have more locks
68 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
69 * down each lock to the point where they're actually slower than just
71 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
73 static DEFINE_SPINLOCK(nbd_lock);
76 static const char *ioctl_cmd_to_ascii(int cmd)
79 case NBD_SET_SOCK: return "set-sock";
80 case NBD_SET_BLKSIZE: return "set-blksize";
81 case NBD_SET_SIZE: return "set-size";
82 case NBD_DO_IT: return "do-it";
83 case NBD_CLEAR_SOCK: return "clear-sock";
84 case NBD_CLEAR_QUE: return "clear-que";
85 case NBD_PRINT_DEBUG: return "print-debug";
86 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
87 case NBD_DISCONNECT: return "disconnect";
88 case BLKROSET: return "set-read-only";
89 case BLKFLSBUF: return "flush-buffer-cache";
94 static const char *nbdcmd_to_ascii(int cmd)
97 case NBD_CMD_READ: return "read";
98 case NBD_CMD_WRITE: return "write";
99 case NBD_CMD_DISC: return "disconnect";
105 static void nbd_end_request(struct request *req)
107 int error = req->errors ? -EIO : 0;
108 struct request_queue *q = req->q;
111 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
112 req, error ? "failed" : "done");
114 spin_lock_irqsave(q->queue_lock, flags);
115 __blk_end_request_all(req, error);
116 spin_unlock_irqrestore(q->queue_lock, flags);
119 static void sock_shutdown(struct nbd_device *lo, int lock)
121 /* Forcibly shutdown the socket causing all listeners
124 * FIXME: This code is duplicated from sys_shutdown, but
125 * there should be a more generic interface rather than
126 * calling socket ops directly here */
128 mutex_lock(&lo->tx_lock);
130 dev_warn(disk_to_dev(lo->disk), "shutting down socket\n");
131 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
135 mutex_unlock(&lo->tx_lock);
138 static void nbd_xmit_timeout(unsigned long arg)
140 struct task_struct *task = (struct task_struct *)arg;
142 printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
143 task->comm, task->pid);
144 force_sig(SIGKILL, task);
148 * Send or receive packet.
150 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
153 struct socket *sock = lo->sock;
157 sigset_t blocked, oldset;
159 if (unlikely(!sock)) {
160 dev_err(disk_to_dev(lo->disk),
161 "Attempted %s on closed socket in sock_xmit\n",
162 (send ? "send" : "recv"));
166 /* Allow interception of SIGKILL only
167 * Don't allow other signals to interrupt the transmission */
168 siginitsetinv(&blocked, sigmask(SIGKILL));
169 sigprocmask(SIG_SETMASK, &blocked, &oldset);
172 sock->sk->sk_allocation = GFP_NOIO;
177 msg.msg_control = NULL;
178 msg.msg_controllen = 0;
179 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
182 struct timer_list ti;
184 if (lo->xmit_timeout) {
186 ti.function = nbd_xmit_timeout;
187 ti.data = (unsigned long)current;
188 ti.expires = jiffies + lo->xmit_timeout;
191 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
192 if (lo->xmit_timeout)
195 result = kernel_recvmsg(sock, &msg, &iov, 1, size,
198 if (signal_pending(current)) {
200 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
201 task_pid_nr(current), current->comm,
202 dequeue_signal_lock(current, ¤t->blocked, &info));
204 sock_shutdown(lo, !send);
210 result = -EPIPE; /* short read */
217 sigprocmask(SIG_SETMASK, &oldset, NULL);
222 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
226 void *kaddr = kmap(bvec->bv_page);
227 result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
228 kunmap(bvec->bv_page);
232 /* always call with the tx_lock held */
233 static int nbd_send_req(struct nbd_device *lo, struct request *req)
236 struct nbd_request request;
237 unsigned long size = blk_rq_bytes(req);
240 request.magic = htonl(NBD_REQUEST_MAGIC);
241 request.type = htonl(nbd_cmd(req));
242 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
243 request.len = htonl(size);
244 memcpy(request.handle, &req, sizeof(req));
246 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
247 lo->disk->disk_name, req,
248 nbdcmd_to_ascii(nbd_cmd(req)),
249 (unsigned long long)blk_rq_pos(req) << 9,
251 result = sock_xmit(lo, 1, &request, sizeof(request),
252 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
254 dev_err(disk_to_dev(lo->disk),
255 "Send control failed (result %d)\n", result);
259 if (nbd_cmd(req) != NBD_CMD_WRITE)
264 struct bio *next = bio->bi_next;
266 struct bio_vec *bvec;
268 bio_for_each_segment(bvec, bio, i) {
269 bool is_last = !next && i == bio->bi_vcnt - 1;
270 int flags = is_last ? 0 : MSG_MORE;
272 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
273 lo->disk->disk_name, req, bvec->bv_len);
274 result = sock_send_bvec(lo, bvec, flags);
276 dev_err(disk_to_dev(lo->disk),
277 "Send data failed (result %d)\n",
282 * The completion might already have come in,
283 * so break for the last one instead of letting
284 * the iterator do it. This prevents use-after-free
298 static struct request *nbd_find_request(struct nbd_device *lo,
299 struct request *xreq)
301 struct request *req, *tmp;
304 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
308 spin_lock(&lo->queue_lock);
309 list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
312 list_del_init(&req->queuelist);
313 spin_unlock(&lo->queue_lock);
316 spin_unlock(&lo->queue_lock);
324 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
327 void *kaddr = kmap(bvec->bv_page);
328 result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
330 kunmap(bvec->bv_page);
334 /* NULL returned = something went wrong, inform userspace */
335 static struct request *nbd_read_stat(struct nbd_device *lo)
338 struct nbd_reply reply;
342 result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
344 dev_err(disk_to_dev(lo->disk),
345 "Receive control failed (result %d)\n", result);
349 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
350 dev_err(disk_to_dev(lo->disk), "Wrong magic (0x%lx)\n",
351 (unsigned long)ntohl(reply.magic));
356 req = nbd_find_request(lo, *(struct request **)reply.handle);
358 result = PTR_ERR(req);
359 if (result != -ENOENT)
362 dev_err(disk_to_dev(lo->disk), "Unexpected reply (%p)\n",
368 if (ntohl(reply.error)) {
369 dev_err(disk_to_dev(lo->disk), "Other side returned error (%d)\n",
375 dprintk(DBG_RX, "%s: request %p: got reply\n",
376 lo->disk->disk_name, req);
377 if (nbd_cmd(req) == NBD_CMD_READ) {
378 struct req_iterator iter;
379 struct bio_vec *bvec;
381 rq_for_each_segment(bvec, req, iter) {
382 result = sock_recv_bvec(lo, bvec);
384 dev_err(disk_to_dev(lo->disk), "Receive data failed (result %d)\n",
389 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
390 lo->disk->disk_name, req, bvec->bv_len);
395 lo->harderror = result;
399 static ssize_t pid_show(struct device *dev,
400 struct device_attribute *attr, char *buf)
402 struct gendisk *disk = dev_to_disk(dev);
404 return sprintf(buf, "%ld\n",
405 (long) ((struct nbd_device *)disk->private_data)->pid);
408 static struct device_attribute pid_attr = {
409 .attr = { .name = "pid", .mode = S_IRUGO},
413 static int nbd_do_it(struct nbd_device *lo)
418 BUG_ON(lo->magic != LO_MAGIC);
420 lo->pid = task_pid_nr(current);
421 ret = device_create_file(disk_to_dev(lo->disk), &pid_attr);
423 dev_err(disk_to_dev(lo->disk), "device_create_file failed!\n");
428 while ((req = nbd_read_stat(lo)) != NULL)
429 nbd_end_request(req);
431 device_remove_file(disk_to_dev(lo->disk), &pid_attr);
436 static void nbd_clear_que(struct nbd_device *lo)
440 BUG_ON(lo->magic != LO_MAGIC);
443 * Because we have set lo->sock to NULL under the tx_lock, all
444 * modifications to the list must have completed by now. For
445 * the same reason, the active_req must be NULL.
447 * As a consequence, we don't need to take the spin lock while
448 * purging the list here.
451 BUG_ON(lo->active_req);
453 while (!list_empty(&lo->queue_head)) {
454 req = list_entry(lo->queue_head.next, struct request,
456 list_del_init(&req->queuelist);
458 nbd_end_request(req);
461 while (!list_empty(&lo->waiting_queue)) {
462 req = list_entry(lo->waiting_queue.next, struct request,
464 list_del_init(&req->queuelist);
466 nbd_end_request(req);
471 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
473 if (req->cmd_type != REQ_TYPE_FS)
476 nbd_cmd(req) = NBD_CMD_READ;
477 if (rq_data_dir(req) == WRITE) {
478 nbd_cmd(req) = NBD_CMD_WRITE;
479 if (lo->flags & NBD_READ_ONLY) {
480 dev_err(disk_to_dev(lo->disk),
481 "Write on read-only\n");
488 mutex_lock(&lo->tx_lock);
489 if (unlikely(!lo->sock)) {
490 mutex_unlock(&lo->tx_lock);
491 dev_err(disk_to_dev(lo->disk),
492 "Attempted send on closed socket\n");
496 lo->active_req = req;
498 if (nbd_send_req(lo, req) != 0) {
499 dev_err(disk_to_dev(lo->disk), "Request send failed\n");
501 nbd_end_request(req);
503 spin_lock(&lo->queue_lock);
504 list_add(&req->queuelist, &lo->queue_head);
505 spin_unlock(&lo->queue_lock);
508 lo->active_req = NULL;
509 mutex_unlock(&lo->tx_lock);
510 wake_up_all(&lo->active_wq);
516 nbd_end_request(req);
519 static int nbd_thread(void *data)
521 struct nbd_device *lo = data;
524 set_user_nice(current, -20);
525 while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
526 /* wait for something to do */
527 wait_event_interruptible(lo->waiting_wq,
528 kthread_should_stop() ||
529 !list_empty(&lo->waiting_queue));
531 /* extract request */
532 if (list_empty(&lo->waiting_queue))
535 spin_lock_irq(&lo->queue_lock);
536 req = list_entry(lo->waiting_queue.next, struct request,
538 list_del_init(&req->queuelist);
539 spin_unlock_irq(&lo->queue_lock);
542 nbd_handle_req(lo, req);
548 * We always wait for result of write, for now. It would be nice to make it optional
550 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
551 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
554 static void do_nbd_request(struct request_queue *q)
558 while ((req = blk_fetch_request(q)) != NULL) {
559 struct nbd_device *lo;
561 spin_unlock_irq(q->queue_lock);
563 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
564 req->rq_disk->disk_name, req, req->cmd_type);
566 lo = req->rq_disk->private_data;
568 BUG_ON(lo->magic != LO_MAGIC);
570 if (unlikely(!lo->sock)) {
571 dev_err(disk_to_dev(lo->disk),
572 "Attempted send on closed socket\n");
574 nbd_end_request(req);
575 spin_lock_irq(q->queue_lock);
579 spin_lock_irq(&lo->queue_lock);
580 list_add_tail(&req->queuelist, &lo->waiting_queue);
581 spin_unlock_irq(&lo->queue_lock);
583 wake_up(&lo->waiting_wq);
585 spin_lock_irq(q->queue_lock);
589 /* Must be called with tx_lock held */
591 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
592 unsigned int cmd, unsigned long arg)
595 case NBD_DISCONNECT: {
598 dev_info(disk_to_dev(lo->disk), "NBD_DISCONNECT\n");
602 mutex_unlock(&lo->tx_lock);
604 mutex_lock(&lo->tx_lock);
605 blk_rq_init(NULL, &sreq);
606 sreq.cmd_type = REQ_TYPE_SPECIAL;
607 nbd_cmd(&sreq) = NBD_CMD_DISC;
609 /* Check again after getting mutex back. */
615 nbd_send_req(lo, &sreq);
619 case NBD_CLEAR_SOCK: {
626 BUG_ON(!list_empty(&lo->queue_head));
627 BUG_ON(!list_empty(&lo->waiting_queue));
640 struct inode *inode = file->f_path.dentry->d_inode;
641 if (S_ISSOCK(inode->i_mode)) {
643 lo->sock = SOCKET_I(inode);
645 bdev->bd_invalidated = 1;
646 lo->disconnect = 0; /* we're connected now */
655 case NBD_SET_BLKSIZE:
657 lo->bytesize &= ~(lo->blksize-1);
658 bdev->bd_inode->i_size = lo->bytesize;
659 set_blocksize(bdev, lo->blksize);
660 set_capacity(lo->disk, lo->bytesize >> 9);
664 lo->bytesize = arg & ~(lo->blksize-1);
665 bdev->bd_inode->i_size = lo->bytesize;
666 set_blocksize(bdev, lo->blksize);
667 set_capacity(lo->disk, lo->bytesize >> 9);
670 case NBD_SET_TIMEOUT:
671 lo->xmit_timeout = arg * HZ;
674 case NBD_SET_SIZE_BLOCKS:
675 lo->bytesize = ((u64) arg) * lo->blksize;
676 bdev->bd_inode->i_size = lo->bytesize;
677 set_blocksize(bdev, lo->blksize);
678 set_capacity(lo->disk, lo->bytesize >> 9);
682 struct task_struct *thread;
691 mutex_unlock(&lo->tx_lock);
693 thread = kthread_create(nbd_thread, lo, "%s",
694 lo->disk->disk_name);
695 if (IS_ERR(thread)) {
696 mutex_lock(&lo->tx_lock);
697 return PTR_ERR(thread);
699 wake_up_process(thread);
700 error = nbd_do_it(lo);
701 kthread_stop(thread);
703 mutex_lock(&lo->tx_lock);
706 sock_shutdown(lo, 0);
710 dev_warn(disk_to_dev(lo->disk), "queue cleared\n");
715 bdev->bd_inode->i_size = 0;
716 set_capacity(lo->disk, 0);
718 ioctl_by_bdev(bdev, BLKRRPART, 0);
719 if (lo->disconnect) /* user requested, ignore socket errors */
721 return lo->harderror;
726 * This is for compatibility only. The queue is always cleared
727 * by NBD_DO_IT or NBD_CLEAR_SOCK.
729 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
732 case NBD_PRINT_DEBUG:
733 dev_info(disk_to_dev(lo->disk),
734 "next = %p, prev = %p, head = %p\n",
735 lo->queue_head.next, lo->queue_head.prev,
742 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
743 unsigned int cmd, unsigned long arg)
745 struct nbd_device *lo = bdev->bd_disk->private_data;
748 if (!capable(CAP_SYS_ADMIN))
751 BUG_ON(lo->magic != LO_MAGIC);
753 /* Anyone capable of this syscall can do *real bad* things */
754 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
755 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
757 mutex_lock(&lo->tx_lock);
758 error = __nbd_ioctl(bdev, lo, cmd, arg);
759 mutex_unlock(&lo->tx_lock);
764 static const struct block_device_operations nbd_fops =
766 .owner = THIS_MODULE,
771 * And here should be modules and kernel interface
772 * (Just smiley confuses emacs :-)
775 static int __init nbd_init(void)
781 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
784 printk(KERN_ERR "nbd: max_part must be >= 0\n");
790 part_shift = fls(max_part);
793 * Adjust max_part according to part_shift as it is exported
794 * to user space so that user can know the max number of
795 * partition kernel should be able to manage.
797 * Note that -1 is required because partition 0 is reserved
798 * for the whole disk.
800 max_part = (1UL << part_shift) - 1;
803 if ((1UL << part_shift) > DISK_MAX_PARTS)
806 if (nbds_max > 1UL << (MINORBITS - part_shift))
809 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
813 for (i = 0; i < nbds_max; i++) {
814 struct gendisk *disk = alloc_disk(1 << part_shift);
817 nbd_dev[i].disk = disk;
819 * The new linux 2.5 block layer implementation requires
820 * every gendisk to have its very own request_queue struct.
821 * These structs are big so we dynamically allocate them.
823 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
829 * Tell the block layer that we are not a rotational device
831 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
834 if (register_blkdev(NBD_MAJOR, "nbd")) {
839 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
840 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
842 for (i = 0; i < nbds_max; i++) {
843 struct gendisk *disk = nbd_dev[i].disk;
844 nbd_dev[i].file = NULL;
845 nbd_dev[i].magic = LO_MAGIC;
846 nbd_dev[i].flags = 0;
847 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
848 spin_lock_init(&nbd_dev[i].queue_lock);
849 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
850 mutex_init(&nbd_dev[i].tx_lock);
851 init_waitqueue_head(&nbd_dev[i].active_wq);
852 init_waitqueue_head(&nbd_dev[i].waiting_wq);
853 nbd_dev[i].blksize = 1024;
854 nbd_dev[i].bytesize = 0;
855 disk->major = NBD_MAJOR;
856 disk->first_minor = i << part_shift;
857 disk->fops = &nbd_fops;
858 disk->private_data = &nbd_dev[i];
859 sprintf(disk->disk_name, "nbd%d", i);
860 set_capacity(disk, 0);
867 blk_cleanup_queue(nbd_dev[i].disk->queue);
868 put_disk(nbd_dev[i].disk);
874 static void __exit nbd_cleanup(void)
877 for (i = 0; i < nbds_max; i++) {
878 struct gendisk *disk = nbd_dev[i].disk;
879 nbd_dev[i].magic = 0;
882 blk_cleanup_queue(disk->queue);
886 unregister_blkdev(NBD_MAJOR, "nbd");
888 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
891 module_init(nbd_init);
892 module_exit(nbd_cleanup);
894 MODULE_DESCRIPTION("Network Block Device");
895 MODULE_LICENSE("GPL");
897 module_param(nbds_max, int, 0444);
898 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
899 module_param(max_part, int, 0444);
900 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
902 module_param(debugflags, int, 0644);
903 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");