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>
35 #include <linux/types.h>
37 #include <asm/uaccess.h>
38 #include <asm/types.h>
40 #include <linux/nbd.h>
44 struct socket * sock; /* If == NULL, device is not ready, yet */
47 spinlock_t queue_lock;
48 struct list_head queue_head; /* Requests waiting result */
49 struct request *active_req;
50 wait_queue_head_t active_wq;
51 struct list_head waiting_queue; /* Requests to be sent */
52 wait_queue_head_t waiting_wq;
59 int disconnect; /* a disconnect has been requested by user */
61 struct timer_list timeout_timer;
62 struct task_struct *task_recv;
63 struct task_struct *task_send;
66 #define NBD_MAGIC 0x68797548
68 static unsigned int nbds_max = 16;
69 static struct nbd_device *nbd_dev;
73 * Use just one lock (or at most 1 per NIC). Two arguments for this:
74 * 1. Each NIC is essentially a synchronization point for all servers
75 * accessed through that NIC so there's no need to have more locks
77 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
78 * down each lock to the point where they're actually slower than just
80 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
82 static DEFINE_SPINLOCK(nbd_lock);
84 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
86 return disk_to_dev(nbd->disk);
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";
95 case NBD_CMD_FLUSH: return "flush";
96 case NBD_CMD_TRIM: return "trim/discard";
101 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
103 int error = req->errors ? -EIO : 0;
104 struct request_queue *q = req->q;
107 dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
108 error ? "failed" : "done");
110 spin_lock_irqsave(q->queue_lock, flags);
111 __blk_end_request_all(req, error);
112 spin_unlock_irqrestore(q->queue_lock, flags);
116 * Forcibly shutdown the socket causing all listeners to error
118 static void sock_shutdown(struct nbd_device *nbd)
123 dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
124 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
126 del_timer_sync(&nbd->timeout_timer);
129 static void nbd_xmit_timeout(unsigned long arg)
131 struct nbd_device *nbd = (struct nbd_device *)arg;
132 struct task_struct *task;
134 if (list_empty(&nbd->queue_head))
139 task = READ_ONCE(nbd->task_recv);
141 force_sig(SIGKILL, task);
143 task = READ_ONCE(nbd->task_send);
145 force_sig(SIGKILL, nbd->task_send);
147 dev_err(nbd_to_dev(nbd), "Connection timed out, killed receiver and sender, shutting down connection\n");
151 * Send or receive packet.
153 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
156 struct socket *sock = nbd->sock;
160 sigset_t blocked, oldset;
161 unsigned long pflags = current->flags;
163 if (unlikely(!sock)) {
164 dev_err(disk_to_dev(nbd->disk),
165 "Attempted %s on closed socket in sock_xmit\n",
166 (send ? "send" : "recv"));
170 /* Allow interception of SIGKILL only
171 * Don't allow other signals to interrupt the transmission */
172 siginitsetinv(&blocked, sigmask(SIGKILL));
173 sigprocmask(SIG_SETMASK, &blocked, &oldset);
175 current->flags |= PF_MEMALLOC;
177 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
182 msg.msg_control = NULL;
183 msg.msg_controllen = 0;
184 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
187 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
189 result = kernel_recvmsg(sock, &msg, &iov, 1, size,
194 result = -EPIPE; /* short read */
201 sigprocmask(SIG_SETMASK, &oldset, NULL);
202 tsk_restore_flags(current, pflags, PF_MEMALLOC);
204 if (!send && nbd->xmit_timeout)
205 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
210 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
214 void *kaddr = kmap(bvec->bv_page);
215 result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
216 bvec->bv_len, flags);
217 kunmap(bvec->bv_page);
221 /* always call with the tx_lock held */
222 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
225 struct nbd_request request;
226 unsigned long size = blk_rq_bytes(req);
229 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
231 else if (req->cmd_flags & REQ_DISCARD)
233 else if (req->cmd_flags & REQ_FLUSH)
234 type = NBD_CMD_FLUSH;
235 else if (rq_data_dir(req) == WRITE)
236 type = NBD_CMD_WRITE;
240 memset(&request, 0, sizeof(request));
241 request.magic = htonl(NBD_REQUEST_MAGIC);
242 request.type = htonl(type);
243 if (type != NBD_CMD_FLUSH && type != NBD_CMD_DISC) {
244 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
245 request.len = htonl(size);
247 memcpy(request.handle, &req, sizeof(req));
249 dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
250 req, nbdcmd_to_ascii(type),
251 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
252 result = sock_xmit(nbd, 1, &request, sizeof(request),
253 (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
255 dev_err(disk_to_dev(nbd->disk),
256 "Send control failed (result %d)\n", result);
260 if (type == NBD_CMD_WRITE) {
261 struct req_iterator iter;
264 * we are really probing at internals to determine
265 * whether to set MSG_MORE or not...
267 rq_for_each_segment(bvec, req, iter) {
269 if (!rq_iter_last(bvec, iter))
271 dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
273 result = sock_send_bvec(nbd, &bvec, flags);
275 dev_err(disk_to_dev(nbd->disk),
276 "Send data failed (result %d)\n",
285 static struct request *nbd_find_request(struct nbd_device *nbd,
286 struct request *xreq)
288 struct request *req, *tmp;
291 err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
295 spin_lock(&nbd->queue_lock);
296 list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
299 list_del_init(&req->queuelist);
300 spin_unlock(&nbd->queue_lock);
303 spin_unlock(&nbd->queue_lock);
305 return ERR_PTR(-ENOENT);
308 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
311 void *kaddr = kmap(bvec->bv_page);
312 result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
314 kunmap(bvec->bv_page);
318 /* NULL returned = something went wrong, inform userspace */
319 static struct request *nbd_read_stat(struct nbd_device *nbd)
322 struct nbd_reply reply;
326 result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
328 dev_err(disk_to_dev(nbd->disk),
329 "Receive control failed (result %d)\n", result);
330 return ERR_PTR(result);
333 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
334 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
335 (unsigned long)ntohl(reply.magic));
336 return ERR_PTR(-EPROTO);
339 req = nbd_find_request(nbd, *(struct request **)reply.handle);
341 result = PTR_ERR(req);
342 if (result != -ENOENT)
343 return ERR_PTR(result);
345 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
347 return ERR_PTR(-EBADR);
350 if (ntohl(reply.error)) {
351 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
357 dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
358 if (rq_data_dir(req) != WRITE) {
359 struct req_iterator iter;
362 rq_for_each_segment(bvec, req, iter) {
363 result = sock_recv_bvec(nbd, &bvec);
365 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
370 dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
377 static ssize_t pid_show(struct device *dev,
378 struct device_attribute *attr, char *buf)
380 struct gendisk *disk = dev_to_disk(dev);
381 struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
383 return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
386 static struct device_attribute pid_attr = {
387 .attr = { .name = "pid", .mode = S_IRUGO},
391 static int nbd_do_it(struct nbd_device *nbd)
396 BUG_ON(nbd->magic != NBD_MAGIC);
398 sk_set_memalloc(nbd->sock->sk);
400 nbd->task_recv = current;
402 ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
404 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
405 nbd->task_recv = NULL;
410 req = nbd_read_stat(nbd);
416 nbd_end_request(nbd, req);
419 device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
421 nbd->task_recv = NULL;
423 if (signal_pending(current)) {
426 ret = dequeue_signal_lock(current, ¤t->blocked, &info);
427 dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
428 task_pid_nr(current), current->comm, ret);
429 mutex_lock(&nbd->tx_lock);
431 mutex_unlock(&nbd->tx_lock);
438 static void nbd_clear_que(struct nbd_device *nbd)
442 BUG_ON(nbd->magic != NBD_MAGIC);
445 * Because we have set nbd->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(nbd->active_req);
455 while (!list_empty(&nbd->queue_head)) {
456 req = list_entry(nbd->queue_head.next, struct request,
458 list_del_init(&req->queuelist);
460 nbd_end_request(nbd, req);
463 while (!list_empty(&nbd->waiting_queue)) {
464 req = list_entry(nbd->waiting_queue.next, struct request,
466 list_del_init(&req->queuelist);
468 nbd_end_request(nbd, req);
470 dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
474 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
476 if (req->cmd_type != REQ_TYPE_FS)
479 if (rq_data_dir(req) == WRITE &&
480 (nbd->flags & NBD_FLAG_READ_ONLY)) {
481 dev_err(disk_to_dev(nbd->disk),
482 "Write on read-only\n");
488 mutex_lock(&nbd->tx_lock);
489 if (unlikely(!nbd->sock)) {
490 mutex_unlock(&nbd->tx_lock);
491 dev_err(disk_to_dev(nbd->disk),
492 "Attempted send on closed socket\n");
496 nbd->active_req = req;
498 if (nbd->xmit_timeout && list_empty_careful(&nbd->queue_head))
499 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
501 if (nbd_send_req(nbd, req) != 0) {
502 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
504 nbd_end_request(nbd, req);
506 spin_lock(&nbd->queue_lock);
507 list_add_tail(&req->queuelist, &nbd->queue_head);
508 spin_unlock(&nbd->queue_lock);
511 nbd->active_req = NULL;
512 mutex_unlock(&nbd->tx_lock);
513 wake_up_all(&nbd->active_wq);
519 nbd_end_request(nbd, req);
522 static int nbd_thread(void *data)
524 struct nbd_device *nbd = data;
527 nbd->task_send = current;
529 set_user_nice(current, MIN_NICE);
530 while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
531 /* wait for something to do */
532 wait_event_interruptible(nbd->waiting_wq,
533 kthread_should_stop() ||
534 !list_empty(&nbd->waiting_queue));
536 if (signal_pending(current)) {
540 ret = dequeue_signal_lock(current, ¤t->blocked,
542 dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
543 task_pid_nr(current), current->comm, ret);
544 mutex_lock(&nbd->tx_lock);
546 mutex_unlock(&nbd->tx_lock);
550 /* extract request */
551 if (list_empty(&nbd->waiting_queue))
554 spin_lock_irq(&nbd->queue_lock);
555 req = list_entry(nbd->waiting_queue.next, struct request,
557 list_del_init(&req->queuelist);
558 spin_unlock_irq(&nbd->queue_lock);
561 nbd_handle_req(nbd, req);
564 nbd->task_send = NULL;
570 * We always wait for result of write, for now. It would be nice to make it optional
572 * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
573 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
576 static void do_nbd_request(struct request_queue *q)
577 __releases(q->queue_lock) __acquires(q->queue_lock)
581 while ((req = blk_fetch_request(q)) != NULL) {
582 struct nbd_device *nbd;
584 spin_unlock_irq(q->queue_lock);
586 nbd = req->rq_disk->private_data;
588 BUG_ON(nbd->magic != NBD_MAGIC);
590 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
593 if (unlikely(!nbd->sock)) {
594 dev_err(disk_to_dev(nbd->disk),
595 "Attempted send on closed socket\n");
597 nbd_end_request(nbd, req);
598 spin_lock_irq(q->queue_lock);
602 spin_lock_irq(&nbd->queue_lock);
603 list_add_tail(&req->queuelist, &nbd->waiting_queue);
604 spin_unlock_irq(&nbd->queue_lock);
606 wake_up(&nbd->waiting_wq);
608 spin_lock_irq(q->queue_lock);
612 /* Must be called with tx_lock held */
614 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
615 unsigned int cmd, unsigned long arg)
618 case NBD_DISCONNECT: {
621 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
625 mutex_unlock(&nbd->tx_lock);
627 mutex_lock(&nbd->tx_lock);
628 blk_rq_init(NULL, &sreq);
629 sreq.cmd_type = REQ_TYPE_DRV_PRIV;
631 /* Check again after getting mutex back. */
637 nbd_send_req(nbd, &sreq);
641 case NBD_CLEAR_SOCK: {
642 struct socket *sock = nbd->sock;
645 BUG_ON(!list_empty(&nbd->queue_head));
646 BUG_ON(!list_empty(&nbd->waiting_queue));
658 sock = sockfd_lookup(arg, &err);
662 bdev->bd_invalidated = 1;
663 nbd->disconnect = 0; /* we're connected now */
669 case NBD_SET_BLKSIZE:
671 nbd->bytesize &= ~(nbd->blksize-1);
672 bdev->bd_inode->i_size = nbd->bytesize;
673 set_blocksize(bdev, nbd->blksize);
674 set_capacity(nbd->disk, nbd->bytesize >> 9);
678 nbd->bytesize = arg & ~(nbd->blksize-1);
679 bdev->bd_inode->i_size = nbd->bytesize;
680 set_blocksize(bdev, nbd->blksize);
681 set_capacity(nbd->disk, nbd->bytesize >> 9);
684 case NBD_SET_TIMEOUT:
685 nbd->xmit_timeout = arg * HZ;
687 mod_timer(&nbd->timeout_timer,
688 jiffies + nbd->xmit_timeout);
690 del_timer_sync(&nbd->timeout_timer);
698 case NBD_SET_SIZE_BLOCKS:
699 nbd->bytesize = ((u64) arg) * nbd->blksize;
700 bdev->bd_inode->i_size = nbd->bytesize;
701 set_blocksize(bdev, nbd->blksize);
702 set_capacity(nbd->disk, nbd->bytesize >> 9);
706 struct task_struct *thread;
715 mutex_unlock(&nbd->tx_lock);
717 if (nbd->flags & NBD_FLAG_READ_ONLY)
718 set_device_ro(bdev, true);
719 if (nbd->flags & NBD_FLAG_SEND_TRIM)
720 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
722 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
723 blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
725 blk_queue_flush(nbd->disk->queue, 0);
727 thread = kthread_run(nbd_thread, nbd, "%s",
728 nbd->disk->disk_name);
729 if (IS_ERR(thread)) {
730 mutex_lock(&nbd->tx_lock);
731 return PTR_ERR(thread);
734 error = nbd_do_it(nbd);
735 kthread_stop(thread);
737 mutex_lock(&nbd->tx_lock);
744 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
745 set_device_ro(bdev, false);
750 bdev->bd_inode->i_size = 0;
751 set_capacity(nbd->disk, 0);
753 blkdev_reread_part(bdev);
754 if (nbd->disconnect) /* user requested, ignore socket errors */
761 * This is for compatibility only. The queue is always cleared
762 * by NBD_DO_IT or NBD_CLEAR_SOCK.
766 case NBD_PRINT_DEBUG:
767 dev_info(disk_to_dev(nbd->disk),
768 "next = %p, prev = %p, head = %p\n",
769 nbd->queue_head.next, nbd->queue_head.prev,
776 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
777 unsigned int cmd, unsigned long arg)
779 struct nbd_device *nbd = bdev->bd_disk->private_data;
782 if (!capable(CAP_SYS_ADMIN))
785 BUG_ON(nbd->magic != NBD_MAGIC);
787 mutex_lock(&nbd->tx_lock);
788 error = __nbd_ioctl(bdev, nbd, cmd, arg);
789 mutex_unlock(&nbd->tx_lock);
794 static const struct block_device_operations nbd_fops =
796 .owner = THIS_MODULE,
801 * And here should be modules and kernel interface
802 * (Just smiley confuses emacs :-)
805 static int __init nbd_init(void)
811 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
814 printk(KERN_ERR "nbd: max_part must be >= 0\n");
820 part_shift = fls(max_part);
823 * Adjust max_part according to part_shift as it is exported
824 * to user space so that user can know the max number of
825 * partition kernel should be able to manage.
827 * Note that -1 is required because partition 0 is reserved
828 * for the whole disk.
830 max_part = (1UL << part_shift) - 1;
833 if ((1UL << part_shift) > DISK_MAX_PARTS)
836 if (nbds_max > 1UL << (MINORBITS - part_shift))
839 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
843 for (i = 0; i < nbds_max; i++) {
844 struct gendisk *disk = alloc_disk(1 << part_shift);
847 nbd_dev[i].disk = disk;
849 * The new linux 2.5 block layer implementation requires
850 * every gendisk to have its very own request_queue struct.
851 * These structs are big so we dynamically allocate them.
853 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
859 * Tell the block layer that we are not a rotational device
861 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
862 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
863 disk->queue->limits.discard_granularity = 512;
864 blk_queue_max_discard_sectors(disk->queue, UINT_MAX);
865 disk->queue->limits.discard_zeroes_data = 0;
866 blk_queue_max_hw_sectors(disk->queue, 65536);
867 disk->queue->limits.max_sectors = 256;
870 if (register_blkdev(NBD_MAJOR, "nbd")) {
875 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
877 for (i = 0; i < nbds_max; i++) {
878 struct gendisk *disk = nbd_dev[i].disk;
879 nbd_dev[i].magic = NBD_MAGIC;
880 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
881 spin_lock_init(&nbd_dev[i].queue_lock);
882 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
883 mutex_init(&nbd_dev[i].tx_lock);
884 init_timer(&nbd_dev[i].timeout_timer);
885 nbd_dev[i].timeout_timer.function = nbd_xmit_timeout;
886 nbd_dev[i].timeout_timer.data = (unsigned long)&nbd_dev[i];
887 init_waitqueue_head(&nbd_dev[i].active_wq);
888 init_waitqueue_head(&nbd_dev[i].waiting_wq);
889 nbd_dev[i].blksize = 1024;
890 nbd_dev[i].bytesize = 0;
891 disk->major = NBD_MAJOR;
892 disk->first_minor = i << part_shift;
893 disk->fops = &nbd_fops;
894 disk->private_data = &nbd_dev[i];
895 sprintf(disk->disk_name, "nbd%d", i);
896 set_capacity(disk, 0);
903 blk_cleanup_queue(nbd_dev[i].disk->queue);
904 put_disk(nbd_dev[i].disk);
910 static void __exit nbd_cleanup(void)
913 for (i = 0; i < nbds_max; i++) {
914 struct gendisk *disk = nbd_dev[i].disk;
915 nbd_dev[i].magic = 0;
918 blk_cleanup_queue(disk->queue);
922 unregister_blkdev(NBD_MAJOR, "nbd");
924 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
927 module_init(nbd_init);
928 module_exit(nbd_cleanup);
930 MODULE_DESCRIPTION("Network Block Device");
931 MODULE_LICENSE("GPL");
933 module_param(nbds_max, int, 0444);
934 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
935 module_param(max_part, int, 0444);
936 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");