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 int harderror; /* Code of hard error */
45 struct socket * sock; /* If == NULL, device is not ready, yet */
48 spinlock_t queue_lock;
49 struct list_head queue_head; /* Requests waiting result */
50 struct request *active_req;
51 wait_queue_head_t active_wq;
52 struct list_head waiting_queue; /* Requests to be sent */
53 wait_queue_head_t waiting_wq;
59 pid_t pid; /* pid of nbd-client, if attached */
61 int disconnect; /* a disconnect has been requested by user */
64 #define NBD_MAGIC 0x68797548
66 static unsigned int nbds_max = 16;
67 static struct nbd_device *nbd_dev;
71 * Use just one lock (or at most 1 per NIC). Two arguments for this:
72 * 1. Each NIC is essentially a synchronization point for all servers
73 * accessed through that NIC so there's no need to have more locks
75 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
76 * down each lock to the point where they're actually slower than just
78 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
80 static DEFINE_SPINLOCK(nbd_lock);
82 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
84 return disk_to_dev(nbd->disk);
87 static const char *nbdcmd_to_ascii(int cmd)
90 case NBD_CMD_READ: return "read";
91 case NBD_CMD_WRITE: return "write";
92 case NBD_CMD_DISC: return "disconnect";
93 case NBD_CMD_FLUSH: return "flush";
94 case NBD_CMD_TRIM: return "trim/discard";
99 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
101 int error = req->errors ? -EIO : 0;
102 struct request_queue *q = req->q;
105 dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
106 error ? "failed" : "done");
108 spin_lock_irqsave(q->queue_lock, flags);
109 __blk_end_request_all(req, error);
110 spin_unlock_irqrestore(q->queue_lock, flags);
114 * Forcibly shutdown the socket causing all listeners to error
116 static void sock_shutdown(struct nbd_device *nbd, int lock)
119 mutex_lock(&nbd->tx_lock);
121 dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
122 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
126 mutex_unlock(&nbd->tx_lock);
129 static void nbd_xmit_timeout(unsigned long arg)
131 struct task_struct *task = (struct task_struct *)arg;
133 printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
134 task->comm, task->pid);
135 force_sig(SIGKILL, task);
139 * Send or receive packet.
141 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
144 struct socket *sock = nbd->sock;
148 sigset_t blocked, oldset;
149 unsigned long pflags = current->flags;
151 if (unlikely(!sock)) {
152 dev_err(disk_to_dev(nbd->disk),
153 "Attempted %s on closed socket in sock_xmit\n",
154 (send ? "send" : "recv"));
158 /* Allow interception of SIGKILL only
159 * Don't allow other signals to interrupt the transmission */
160 siginitsetinv(&blocked, sigmask(SIGKILL));
161 sigprocmask(SIG_SETMASK, &blocked, &oldset);
163 current->flags |= PF_MEMALLOC;
165 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
170 msg.msg_control = NULL;
171 msg.msg_controllen = 0;
172 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
175 struct timer_list ti;
177 if (nbd->xmit_timeout) {
179 ti.function = nbd_xmit_timeout;
180 ti.data = (unsigned long)current;
181 ti.expires = jiffies + nbd->xmit_timeout;
184 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
185 if (nbd->xmit_timeout)
188 result = kernel_recvmsg(sock, &msg, &iov, 1, size,
191 if (signal_pending(current)) {
193 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
194 task_pid_nr(current), current->comm,
195 dequeue_signal_lock(current, ¤t->blocked, &info));
197 sock_shutdown(nbd, !send);
203 result = -EPIPE; /* short read */
210 sigprocmask(SIG_SETMASK, &oldset, NULL);
211 tsk_restore_flags(current, pflags, PF_MEMALLOC);
216 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
220 void *kaddr = kmap(bvec->bv_page);
221 result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
222 bvec->bv_len, flags);
223 kunmap(bvec->bv_page);
227 /* always call with the tx_lock held */
228 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
231 struct nbd_request request;
232 unsigned long size = blk_rq_bytes(req);
235 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
237 else if (req->cmd_flags & REQ_DISCARD)
239 else if (req->cmd_flags & REQ_FLUSH)
240 type = NBD_CMD_FLUSH;
241 else if (rq_data_dir(req) == WRITE)
242 type = NBD_CMD_WRITE;
246 memset(&request, 0, sizeof(request));
247 request.magic = htonl(NBD_REQUEST_MAGIC);
248 request.type = htonl(type);
249 if (type != NBD_CMD_FLUSH && type != NBD_CMD_DISC) {
250 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
251 request.len = htonl(size);
253 memcpy(request.handle, &req, sizeof(req));
255 dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
256 req, nbdcmd_to_ascii(type),
257 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
258 result = sock_xmit(nbd, 1, &request, sizeof(request),
259 (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
261 dev_err(disk_to_dev(nbd->disk),
262 "Send control failed (result %d)\n", result);
266 if (type == NBD_CMD_WRITE) {
267 struct req_iterator iter;
270 * we are really probing at internals to determine
271 * whether to set MSG_MORE or not...
273 rq_for_each_segment(bvec, req, iter) {
275 if (!rq_iter_last(bvec, iter))
277 dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
279 result = sock_send_bvec(nbd, &bvec, flags);
281 dev_err(disk_to_dev(nbd->disk),
282 "Send data failed (result %d)\n",
291 static struct request *nbd_find_request(struct nbd_device *nbd,
292 struct request *xreq)
294 struct request *req, *tmp;
297 err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
301 spin_lock(&nbd->queue_lock);
302 list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
305 list_del_init(&req->queuelist);
306 spin_unlock(&nbd->queue_lock);
309 spin_unlock(&nbd->queue_lock);
311 return ERR_PTR(-ENOENT);
314 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
317 void *kaddr = kmap(bvec->bv_page);
318 result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
320 kunmap(bvec->bv_page);
324 /* NULL returned = something went wrong, inform userspace */
325 static struct request *nbd_read_stat(struct nbd_device *nbd)
328 struct nbd_reply reply;
332 result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
334 dev_err(disk_to_dev(nbd->disk),
335 "Receive control failed (result %d)\n", result);
339 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
340 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
341 (unsigned long)ntohl(reply.magic));
346 req = nbd_find_request(nbd, *(struct request **)reply.handle);
348 result = PTR_ERR(req);
349 if (result != -ENOENT)
352 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
358 if (ntohl(reply.error)) {
359 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
365 dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
366 if (rq_data_dir(req) != WRITE) {
367 struct req_iterator iter;
370 rq_for_each_segment(bvec, req, iter) {
371 result = sock_recv_bvec(nbd, &bvec);
373 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
378 dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
384 nbd->harderror = result;
388 static ssize_t pid_show(struct device *dev,
389 struct device_attribute *attr, char *buf)
391 struct gendisk *disk = dev_to_disk(dev);
393 return sprintf(buf, "%ld\n",
394 (long) ((struct nbd_device *)disk->private_data)->pid);
397 static struct device_attribute pid_attr = {
398 .attr = { .name = "pid", .mode = S_IRUGO},
402 static int nbd_do_it(struct nbd_device *nbd)
407 BUG_ON(nbd->magic != NBD_MAGIC);
409 sk_set_memalloc(nbd->sock->sk);
410 nbd->pid = task_pid_nr(current);
411 ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
413 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
418 while ((req = nbd_read_stat(nbd)) != NULL)
419 nbd_end_request(nbd, req);
421 device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
426 static void nbd_clear_que(struct nbd_device *nbd)
430 BUG_ON(nbd->magic != NBD_MAGIC);
433 * Because we have set nbd->sock to NULL under the tx_lock, all
434 * modifications to the list must have completed by now. For
435 * the same reason, the active_req must be NULL.
437 * As a consequence, we don't need to take the spin lock while
438 * purging the list here.
441 BUG_ON(nbd->active_req);
443 while (!list_empty(&nbd->queue_head)) {
444 req = list_entry(nbd->queue_head.next, struct request,
446 list_del_init(&req->queuelist);
448 nbd_end_request(nbd, req);
451 while (!list_empty(&nbd->waiting_queue)) {
452 req = list_entry(nbd->waiting_queue.next, struct request,
454 list_del_init(&req->queuelist);
456 nbd_end_request(nbd, req);
461 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
463 if (req->cmd_type != REQ_TYPE_FS)
466 if (rq_data_dir(req) == WRITE &&
467 (nbd->flags & NBD_FLAG_READ_ONLY)) {
468 dev_err(disk_to_dev(nbd->disk),
469 "Write on read-only\n");
475 mutex_lock(&nbd->tx_lock);
476 if (unlikely(!nbd->sock)) {
477 mutex_unlock(&nbd->tx_lock);
478 dev_err(disk_to_dev(nbd->disk),
479 "Attempted send on closed socket\n");
483 nbd->active_req = req;
485 if (nbd_send_req(nbd, req) != 0) {
486 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
488 nbd_end_request(nbd, req);
490 spin_lock(&nbd->queue_lock);
491 list_add_tail(&req->queuelist, &nbd->queue_head);
492 spin_unlock(&nbd->queue_lock);
495 nbd->active_req = NULL;
496 mutex_unlock(&nbd->tx_lock);
497 wake_up_all(&nbd->active_wq);
503 nbd_end_request(nbd, req);
506 static int nbd_thread(void *data)
508 struct nbd_device *nbd = data;
511 set_user_nice(current, MIN_NICE);
512 while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
513 /* wait for something to do */
514 wait_event_interruptible(nbd->waiting_wq,
515 kthread_should_stop() ||
516 !list_empty(&nbd->waiting_queue));
518 /* extract request */
519 if (list_empty(&nbd->waiting_queue))
522 spin_lock_irq(&nbd->queue_lock);
523 req = list_entry(nbd->waiting_queue.next, struct request,
525 list_del_init(&req->queuelist);
526 spin_unlock_irq(&nbd->queue_lock);
529 nbd_handle_req(nbd, req);
535 * We always wait for result of write, for now. It would be nice to make it optional
537 * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
538 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
541 static void do_nbd_request(struct request_queue *q)
542 __releases(q->queue_lock) __acquires(q->queue_lock)
546 while ((req = blk_fetch_request(q)) != NULL) {
547 struct nbd_device *nbd;
549 spin_unlock_irq(q->queue_lock);
551 nbd = req->rq_disk->private_data;
553 BUG_ON(nbd->magic != NBD_MAGIC);
555 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
558 if (unlikely(!nbd->sock)) {
559 dev_err(disk_to_dev(nbd->disk),
560 "Attempted send on closed socket\n");
562 nbd_end_request(nbd, req);
563 spin_lock_irq(q->queue_lock);
567 spin_lock_irq(&nbd->queue_lock);
568 list_add_tail(&req->queuelist, &nbd->waiting_queue);
569 spin_unlock_irq(&nbd->queue_lock);
571 wake_up(&nbd->waiting_wq);
573 spin_lock_irq(q->queue_lock);
577 /* Must be called with tx_lock held */
579 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
580 unsigned int cmd, unsigned long arg)
583 case NBD_DISCONNECT: {
586 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
590 mutex_unlock(&nbd->tx_lock);
592 mutex_lock(&nbd->tx_lock);
593 blk_rq_init(NULL, &sreq);
594 sreq.cmd_type = REQ_TYPE_DRV_PRIV;
596 /* Check again after getting mutex back. */
602 nbd_send_req(nbd, &sreq);
606 case NBD_CLEAR_SOCK: {
607 struct socket *sock = nbd->sock;
610 BUG_ON(!list_empty(&nbd->queue_head));
611 BUG_ON(!list_empty(&nbd->waiting_queue));
623 sock = sockfd_lookup(arg, &err);
627 bdev->bd_invalidated = 1;
628 nbd->disconnect = 0; /* we're connected now */
634 case NBD_SET_BLKSIZE:
636 nbd->bytesize &= ~(nbd->blksize-1);
637 bdev->bd_inode->i_size = nbd->bytesize;
638 set_blocksize(bdev, nbd->blksize);
639 set_capacity(nbd->disk, nbd->bytesize >> 9);
643 nbd->bytesize = arg & ~(nbd->blksize-1);
644 bdev->bd_inode->i_size = nbd->bytesize;
645 set_blocksize(bdev, nbd->blksize);
646 set_capacity(nbd->disk, nbd->bytesize >> 9);
649 case NBD_SET_TIMEOUT:
650 nbd->xmit_timeout = arg * HZ;
657 case NBD_SET_SIZE_BLOCKS:
658 nbd->bytesize = ((u64) arg) * nbd->blksize;
659 bdev->bd_inode->i_size = nbd->bytesize;
660 set_blocksize(bdev, nbd->blksize);
661 set_capacity(nbd->disk, nbd->bytesize >> 9);
665 struct task_struct *thread;
674 mutex_unlock(&nbd->tx_lock);
676 if (nbd->flags & NBD_FLAG_READ_ONLY)
677 set_device_ro(bdev, true);
678 if (nbd->flags & NBD_FLAG_SEND_TRIM)
679 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
681 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
682 blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
684 blk_queue_flush(nbd->disk->queue, 0);
686 thread = kthread_run(nbd_thread, nbd, "%s",
687 nbd->disk->disk_name);
688 if (IS_ERR(thread)) {
689 mutex_lock(&nbd->tx_lock);
690 return PTR_ERR(thread);
693 error = nbd_do_it(nbd);
694 kthread_stop(thread);
696 mutex_lock(&nbd->tx_lock);
699 sock_shutdown(nbd, 0);
703 dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
705 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
706 set_device_ro(bdev, false);
711 bdev->bd_inode->i_size = 0;
712 set_capacity(nbd->disk, 0);
714 blkdev_reread_part(bdev);
715 if (nbd->disconnect) /* user requested, ignore socket errors */
717 return nbd->harderror;
722 * This is for compatibility only. The queue is always cleared
723 * by NBD_DO_IT or NBD_CLEAR_SOCK.
727 case NBD_PRINT_DEBUG:
728 dev_info(disk_to_dev(nbd->disk),
729 "next = %p, prev = %p, head = %p\n",
730 nbd->queue_head.next, nbd->queue_head.prev,
737 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
738 unsigned int cmd, unsigned long arg)
740 struct nbd_device *nbd = bdev->bd_disk->private_data;
743 if (!capable(CAP_SYS_ADMIN))
746 BUG_ON(nbd->magic != NBD_MAGIC);
748 mutex_lock(&nbd->tx_lock);
749 error = __nbd_ioctl(bdev, nbd, cmd, arg);
750 mutex_unlock(&nbd->tx_lock);
755 static const struct block_device_operations nbd_fops =
757 .owner = THIS_MODULE,
762 * And here should be modules and kernel interface
763 * (Just smiley confuses emacs :-)
766 static int __init nbd_init(void)
772 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
775 printk(KERN_ERR "nbd: max_part must be >= 0\n");
781 part_shift = fls(max_part);
784 * Adjust max_part according to part_shift as it is exported
785 * to user space so that user can know the max number of
786 * partition kernel should be able to manage.
788 * Note that -1 is required because partition 0 is reserved
789 * for the whole disk.
791 max_part = (1UL << part_shift) - 1;
794 if ((1UL << part_shift) > DISK_MAX_PARTS)
797 if (nbds_max > 1UL << (MINORBITS - part_shift))
800 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
804 for (i = 0; i < nbds_max; i++) {
805 struct gendisk *disk = alloc_disk(1 << part_shift);
808 nbd_dev[i].disk = disk;
810 * The new linux 2.5 block layer implementation requires
811 * every gendisk to have its very own request_queue struct.
812 * These structs are big so we dynamically allocate them.
814 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
820 * Tell the block layer that we are not a rotational device
822 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
823 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
824 disk->queue->limits.discard_granularity = 512;
825 disk->queue->limits.max_discard_sectors = UINT_MAX;
826 disk->queue->limits.discard_zeroes_data = 0;
827 blk_queue_max_hw_sectors(disk->queue, 65536);
828 disk->queue->limits.max_sectors = 256;
831 if (register_blkdev(NBD_MAJOR, "nbd")) {
836 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
838 for (i = 0; i < nbds_max; i++) {
839 struct gendisk *disk = nbd_dev[i].disk;
840 nbd_dev[i].magic = NBD_MAGIC;
841 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
842 spin_lock_init(&nbd_dev[i].queue_lock);
843 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
844 mutex_init(&nbd_dev[i].tx_lock);
845 init_waitqueue_head(&nbd_dev[i].active_wq);
846 init_waitqueue_head(&nbd_dev[i].waiting_wq);
847 nbd_dev[i].blksize = 1024;
848 nbd_dev[i].bytesize = 0;
849 disk->major = NBD_MAJOR;
850 disk->first_minor = i << part_shift;
851 disk->fops = &nbd_fops;
852 disk->private_data = &nbd_dev[i];
853 sprintf(disk->disk_name, "nbd%d", i);
854 set_capacity(disk, 0);
861 blk_cleanup_queue(nbd_dev[i].disk->queue);
862 put_disk(nbd_dev[i].disk);
868 static void __exit nbd_cleanup(void)
871 for (i = 0; i < nbds_max; i++) {
872 struct gendisk *disk = nbd_dev[i].disk;
873 nbd_dev[i].magic = 0;
876 blk_cleanup_queue(disk->queue);
880 unregister_blkdev(NBD_MAJOR, "nbd");
882 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
885 module_init(nbd_init);
886 module_exit(nbd_cleanup);
888 MODULE_DESCRIPTION("Network Block Device");
889 MODULE_LICENSE("GPL");
891 module_param(nbds_max, int, 0444);
892 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
893 module_param(max_part, int, 0444);
894 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");