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)
265 struct bio *next = bio->bi_next;
267 struct bio_vec *bvec;
269 bio_for_each_segment(bvec, bio, i) {
270 bool is_last = !next && i == bio->bi_vcnt - 1;
274 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
275 lo->disk->disk_name, req, bvec->bv_len);
276 result = sock_send_bvec(lo, bvec, flags);
278 dev_err(disk_to_dev(lo->disk),
279 "Send data failed (result %d)\n",
284 * The completion might already have come in,
285 * so break for the last one instead of letting
286 * the iterator do it. This prevents use-after-free
300 static struct request *nbd_find_request(struct nbd_device *lo,
301 struct request *xreq)
303 struct request *req, *tmp;
306 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
310 spin_lock(&lo->queue_lock);
311 list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
314 list_del_init(&req->queuelist);
315 spin_unlock(&lo->queue_lock);
318 spin_unlock(&lo->queue_lock);
326 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
329 void *kaddr = kmap(bvec->bv_page);
330 result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
332 kunmap(bvec->bv_page);
336 /* NULL returned = something went wrong, inform userspace */
337 static struct request *nbd_read_stat(struct nbd_device *lo)
340 struct nbd_reply reply;
344 result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
346 dev_err(disk_to_dev(lo->disk),
347 "Receive control failed (result %d)\n", result);
351 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
352 dev_err(disk_to_dev(lo->disk), "Wrong magic (0x%lx)\n",
353 (unsigned long)ntohl(reply.magic));
358 req = nbd_find_request(lo, *(struct request **)reply.handle);
360 result = PTR_ERR(req);
361 if (result != -ENOENT)
364 dev_err(disk_to_dev(lo->disk), "Unexpected reply (%p)\n",
370 if (ntohl(reply.error)) {
371 dev_err(disk_to_dev(lo->disk), "Other side returned error (%d)\n",
377 dprintk(DBG_RX, "%s: request %p: got reply\n",
378 lo->disk->disk_name, req);
379 if (nbd_cmd(req) == NBD_CMD_READ) {
380 struct req_iterator iter;
381 struct bio_vec *bvec;
383 rq_for_each_segment(bvec, req, iter) {
384 result = sock_recv_bvec(lo, bvec);
386 dev_err(disk_to_dev(lo->disk), "Receive data failed (result %d)\n",
391 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
392 lo->disk->disk_name, req, bvec->bv_len);
397 lo->harderror = result;
401 static ssize_t pid_show(struct device *dev,
402 struct device_attribute *attr, char *buf)
404 struct gendisk *disk = dev_to_disk(dev);
406 return sprintf(buf, "%ld\n",
407 (long) ((struct nbd_device *)disk->private_data)->pid);
410 static struct device_attribute pid_attr = {
411 .attr = { .name = "pid", .mode = S_IRUGO},
415 static int nbd_do_it(struct nbd_device *lo)
420 BUG_ON(lo->magic != LO_MAGIC);
422 lo->pid = task_pid_nr(current);
423 ret = device_create_file(disk_to_dev(lo->disk), &pid_attr);
425 dev_err(disk_to_dev(lo->disk), "device_create_file failed!\n");
430 while ((req = nbd_read_stat(lo)) != NULL)
431 nbd_end_request(req);
433 device_remove_file(disk_to_dev(lo->disk), &pid_attr);
438 static void nbd_clear_que(struct nbd_device *lo)
442 BUG_ON(lo->magic != LO_MAGIC);
445 * Because we have set lo->sock to NULL under the tx_lock, all
446 * modifications to the list must have completed by now. For
447 * the same reason, the active_req must be NULL.
449 * As a consequence, we don't need to take the spin lock while
450 * purging the list here.
453 BUG_ON(lo->active_req);
455 while (!list_empty(&lo->queue_head)) {
456 req = list_entry(lo->queue_head.next, struct request,
458 list_del_init(&req->queuelist);
460 nbd_end_request(req);
463 while (!list_empty(&lo->waiting_queue)) {
464 req = list_entry(lo->waiting_queue.next, struct request,
466 list_del_init(&req->queuelist);
468 nbd_end_request(req);
473 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
475 if (req->cmd_type != REQ_TYPE_FS)
478 nbd_cmd(req) = NBD_CMD_READ;
479 if (rq_data_dir(req) == WRITE) {
480 nbd_cmd(req) = NBD_CMD_WRITE;
481 if (lo->flags & NBD_READ_ONLY) {
482 dev_err(disk_to_dev(lo->disk),
483 "Write on read-only\n");
490 mutex_lock(&lo->tx_lock);
491 if (unlikely(!lo->sock)) {
492 mutex_unlock(&lo->tx_lock);
493 dev_err(disk_to_dev(lo->disk),
494 "Attempted send on closed socket\n");
498 lo->active_req = req;
500 if (nbd_send_req(lo, req) != 0) {
501 dev_err(disk_to_dev(lo->disk), "Request send failed\n");
503 nbd_end_request(req);
505 spin_lock(&lo->queue_lock);
506 list_add(&req->queuelist, &lo->queue_head);
507 spin_unlock(&lo->queue_lock);
510 lo->active_req = NULL;
511 mutex_unlock(&lo->tx_lock);
512 wake_up_all(&lo->active_wq);
518 nbd_end_request(req);
521 static int nbd_thread(void *data)
523 struct nbd_device *lo = data;
526 set_user_nice(current, -20);
527 while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
528 /* wait for something to do */
529 wait_event_interruptible(lo->waiting_wq,
530 kthread_should_stop() ||
531 !list_empty(&lo->waiting_queue));
533 /* extract request */
534 if (list_empty(&lo->waiting_queue))
537 spin_lock_irq(&lo->queue_lock);
538 req = list_entry(lo->waiting_queue.next, struct request,
540 list_del_init(&req->queuelist);
541 spin_unlock_irq(&lo->queue_lock);
544 nbd_handle_req(lo, req);
550 * We always wait for result of write, for now. It would be nice to make it optional
552 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
553 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
556 static void do_nbd_request(struct request_queue *q)
560 while ((req = blk_fetch_request(q)) != NULL) {
561 struct nbd_device *lo;
563 spin_unlock_irq(q->queue_lock);
565 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
566 req->rq_disk->disk_name, req, req->cmd_type);
568 lo = req->rq_disk->private_data;
570 BUG_ON(lo->magic != LO_MAGIC);
572 if (unlikely(!lo->sock)) {
573 dev_err(disk_to_dev(lo->disk),
574 "Attempted send on closed socket\n");
576 nbd_end_request(req);
577 spin_lock_irq(q->queue_lock);
581 spin_lock_irq(&lo->queue_lock);
582 list_add_tail(&req->queuelist, &lo->waiting_queue);
583 spin_unlock_irq(&lo->queue_lock);
585 wake_up(&lo->waiting_wq);
587 spin_lock_irq(q->queue_lock);
591 /* Must be called with tx_lock held */
593 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
594 unsigned int cmd, unsigned long arg)
597 case NBD_DISCONNECT: {
600 dev_info(disk_to_dev(lo->disk), "NBD_DISCONNECT\n");
604 mutex_unlock(&lo->tx_lock);
606 mutex_lock(&lo->tx_lock);
607 blk_rq_init(NULL, &sreq);
608 sreq.cmd_type = REQ_TYPE_SPECIAL;
609 nbd_cmd(&sreq) = NBD_CMD_DISC;
611 /* Check again after getting mutex back. */
617 nbd_send_req(lo, &sreq);
621 case NBD_CLEAR_SOCK: {
628 BUG_ON(!list_empty(&lo->queue_head));
629 BUG_ON(!list_empty(&lo->waiting_queue));
642 struct inode *inode = file->f_path.dentry->d_inode;
643 if (S_ISSOCK(inode->i_mode)) {
645 lo->sock = SOCKET_I(inode);
647 bdev->bd_invalidated = 1;
648 lo->disconnect = 0; /* we're connected now */
657 case NBD_SET_BLKSIZE:
659 lo->bytesize &= ~(lo->blksize-1);
660 bdev->bd_inode->i_size = lo->bytesize;
661 set_blocksize(bdev, lo->blksize);
662 set_capacity(lo->disk, lo->bytesize >> 9);
666 lo->bytesize = arg & ~(lo->blksize-1);
667 bdev->bd_inode->i_size = lo->bytesize;
668 set_blocksize(bdev, lo->blksize);
669 set_capacity(lo->disk, lo->bytesize >> 9);
672 case NBD_SET_TIMEOUT:
673 lo->xmit_timeout = arg * HZ;
676 case NBD_SET_SIZE_BLOCKS:
677 lo->bytesize = ((u64) arg) * lo->blksize;
678 bdev->bd_inode->i_size = lo->bytesize;
679 set_blocksize(bdev, lo->blksize);
680 set_capacity(lo->disk, lo->bytesize >> 9);
684 struct task_struct *thread;
693 mutex_unlock(&lo->tx_lock);
695 thread = kthread_create(nbd_thread, lo, "%s",
696 lo->disk->disk_name);
697 if (IS_ERR(thread)) {
698 mutex_lock(&lo->tx_lock);
699 return PTR_ERR(thread);
701 wake_up_process(thread);
702 error = nbd_do_it(lo);
703 kthread_stop(thread);
705 mutex_lock(&lo->tx_lock);
708 sock_shutdown(lo, 0);
712 dev_warn(disk_to_dev(lo->disk), "queue cleared\n");
717 bdev->bd_inode->i_size = 0;
718 set_capacity(lo->disk, 0);
720 ioctl_by_bdev(bdev, BLKRRPART, 0);
721 if (lo->disconnect) /* user requested, ignore socket errors */
723 return lo->harderror;
728 * This is for compatibility only. The queue is always cleared
729 * by NBD_DO_IT or NBD_CLEAR_SOCK.
731 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
734 case NBD_PRINT_DEBUG:
735 dev_info(disk_to_dev(lo->disk),
736 "next = %p, prev = %p, head = %p\n",
737 lo->queue_head.next, lo->queue_head.prev,
744 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
745 unsigned int cmd, unsigned long arg)
747 struct nbd_device *lo = bdev->bd_disk->private_data;
750 if (!capable(CAP_SYS_ADMIN))
753 BUG_ON(lo->magic != LO_MAGIC);
755 /* Anyone capable of this syscall can do *real bad* things */
756 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
757 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
759 mutex_lock(&lo->tx_lock);
760 error = __nbd_ioctl(bdev, lo, cmd, arg);
761 mutex_unlock(&lo->tx_lock);
766 static const struct block_device_operations nbd_fops =
768 .owner = THIS_MODULE,
773 * And here should be modules and kernel interface
774 * (Just smiley confuses emacs :-)
777 static int __init nbd_init(void)
783 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
786 printk(KERN_ERR "nbd: max_part must be >= 0\n");
792 part_shift = fls(max_part);
795 * Adjust max_part according to part_shift as it is exported
796 * to user space so that user can know the max number of
797 * partition kernel should be able to manage.
799 * Note that -1 is required because partition 0 is reserved
800 * for the whole disk.
802 max_part = (1UL << part_shift) - 1;
805 if ((1UL << part_shift) > DISK_MAX_PARTS)
808 if (nbds_max > 1UL << (MINORBITS - part_shift))
811 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
815 for (i = 0; i < nbds_max; i++) {
816 struct gendisk *disk = alloc_disk(1 << part_shift);
819 nbd_dev[i].disk = disk;
821 * The new linux 2.5 block layer implementation requires
822 * every gendisk to have its very own request_queue struct.
823 * These structs are big so we dynamically allocate them.
825 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
831 * Tell the block layer that we are not a rotational device
833 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
836 if (register_blkdev(NBD_MAJOR, "nbd")) {
841 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
842 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
844 for (i = 0; i < nbds_max; i++) {
845 struct gendisk *disk = nbd_dev[i].disk;
846 nbd_dev[i].file = NULL;
847 nbd_dev[i].magic = LO_MAGIC;
848 nbd_dev[i].flags = 0;
849 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
850 spin_lock_init(&nbd_dev[i].queue_lock);
851 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
852 mutex_init(&nbd_dev[i].tx_lock);
853 init_waitqueue_head(&nbd_dev[i].active_wq);
854 init_waitqueue_head(&nbd_dev[i].waiting_wq);
855 nbd_dev[i].blksize = 1024;
856 nbd_dev[i].bytesize = 0;
857 disk->major = NBD_MAJOR;
858 disk->first_minor = i << part_shift;
859 disk->fops = &nbd_fops;
860 disk->private_data = &nbd_dev[i];
861 sprintf(disk->disk_name, "nbd%d", i);
862 set_capacity(disk, 0);
869 blk_cleanup_queue(nbd_dev[i].disk->queue);
870 put_disk(nbd_dev[i].disk);
876 static void __exit nbd_cleanup(void)
879 for (i = 0; i < nbds_max; i++) {
880 struct gendisk *disk = nbd_dev[i].disk;
881 nbd_dev[i].magic = 0;
884 blk_cleanup_queue(disk->queue);
888 unregister_blkdev(NBD_MAJOR, "nbd");
890 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
893 module_init(nbd_init);
894 module_exit(nbd_cleanup);
896 MODULE_DESCRIPTION("Network Block Device");
897 MODULE_LICENSE("GPL");
899 module_param(nbds_max, int, 0444);
900 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
901 module_param(max_part, int, 0444);
902 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
904 module_param(debugflags, int, 0644);
905 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");