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);
234 memset(&request, 0, sizeof(request));
235 request.magic = htonl(NBD_REQUEST_MAGIC);
236 request.type = htonl(nbd_cmd(req));
238 if (nbd_cmd(req) != NBD_CMD_FLUSH && nbd_cmd(req) != NBD_CMD_DISC) {
239 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
240 request.len = htonl(size);
242 memcpy(request.handle, &req, sizeof(req));
244 dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
245 req, nbdcmd_to_ascii(nbd_cmd(req)),
246 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
247 result = sock_xmit(nbd, 1, &request, sizeof(request),
248 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
250 dev_err(disk_to_dev(nbd->disk),
251 "Send control failed (result %d)\n", result);
255 if (nbd_cmd(req) == NBD_CMD_WRITE) {
256 struct req_iterator iter;
259 * we are really probing at internals to determine
260 * whether to set MSG_MORE or not...
262 rq_for_each_segment(bvec, req, iter) {
264 if (!rq_iter_last(bvec, iter))
266 dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
268 result = sock_send_bvec(nbd, &bvec, flags);
270 dev_err(disk_to_dev(nbd->disk),
271 "Send data failed (result %d)\n",
283 static struct request *nbd_find_request(struct nbd_device *nbd,
284 struct request *xreq)
286 struct request *req, *tmp;
289 err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
293 spin_lock(&nbd->queue_lock);
294 list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
297 list_del_init(&req->queuelist);
298 spin_unlock(&nbd->queue_lock);
301 spin_unlock(&nbd->queue_lock);
309 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
312 void *kaddr = kmap(bvec->bv_page);
313 result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
315 kunmap(bvec->bv_page);
319 /* NULL returned = something went wrong, inform userspace */
320 static struct request *nbd_read_stat(struct nbd_device *nbd)
323 struct nbd_reply reply;
327 result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
329 dev_err(disk_to_dev(nbd->disk),
330 "Receive control failed (result %d)\n", result);
334 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
335 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
336 (unsigned long)ntohl(reply.magic));
341 req = nbd_find_request(nbd, *(struct request **)reply.handle);
343 result = PTR_ERR(req);
344 if (result != -ENOENT)
347 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
353 if (ntohl(reply.error)) {
354 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
360 dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
361 if (nbd_cmd(req) == NBD_CMD_READ) {
362 struct req_iterator iter;
365 rq_for_each_segment(bvec, req, iter) {
366 result = sock_recv_bvec(nbd, &bvec);
368 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
373 dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
379 nbd->harderror = result;
383 static ssize_t pid_show(struct device *dev,
384 struct device_attribute *attr, char *buf)
386 struct gendisk *disk = dev_to_disk(dev);
388 return sprintf(buf, "%ld\n",
389 (long) ((struct nbd_device *)disk->private_data)->pid);
392 static struct device_attribute pid_attr = {
393 .attr = { .name = "pid", .mode = S_IRUGO},
397 static int nbd_do_it(struct nbd_device *nbd)
402 BUG_ON(nbd->magic != NBD_MAGIC);
404 sk_set_memalloc(nbd->sock->sk);
405 nbd->pid = task_pid_nr(current);
406 ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
408 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
413 while ((req = nbd_read_stat(nbd)) != NULL)
414 nbd_end_request(nbd, req);
416 device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
421 static void nbd_clear_que(struct nbd_device *nbd)
425 BUG_ON(nbd->magic != NBD_MAGIC);
428 * Because we have set nbd->sock to NULL under the tx_lock, all
429 * modifications to the list must have completed by now. For
430 * the same reason, the active_req must be NULL.
432 * As a consequence, we don't need to take the spin lock while
433 * purging the list here.
436 BUG_ON(nbd->active_req);
438 while (!list_empty(&nbd->queue_head)) {
439 req = list_entry(nbd->queue_head.next, struct request,
441 list_del_init(&req->queuelist);
443 nbd_end_request(nbd, req);
446 while (!list_empty(&nbd->waiting_queue)) {
447 req = list_entry(nbd->waiting_queue.next, struct request,
449 list_del_init(&req->queuelist);
451 nbd_end_request(nbd, req);
456 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
458 if (req->cmd_type != REQ_TYPE_FS)
461 nbd_cmd(req) = NBD_CMD_READ;
462 if (rq_data_dir(req) == WRITE) {
463 if ((req->cmd_flags & REQ_DISCARD)) {
464 WARN_ON(!(nbd->flags & NBD_FLAG_SEND_TRIM));
465 nbd_cmd(req) = NBD_CMD_TRIM;
467 nbd_cmd(req) = NBD_CMD_WRITE;
468 if (nbd->flags & NBD_FLAG_READ_ONLY) {
469 dev_err(disk_to_dev(nbd->disk),
470 "Write on read-only\n");
475 if (req->cmd_flags & REQ_FLUSH) {
476 BUG_ON(unlikely(blk_rq_sectors(req)));
477 nbd_cmd(req) = NBD_CMD_FLUSH;
482 mutex_lock(&nbd->tx_lock);
483 if (unlikely(!nbd->sock)) {
484 mutex_unlock(&nbd->tx_lock);
485 dev_err(disk_to_dev(nbd->disk),
486 "Attempted send on closed socket\n");
490 nbd->active_req = req;
492 if (nbd_send_req(nbd, req) != 0) {
493 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
495 nbd_end_request(nbd, req);
497 spin_lock(&nbd->queue_lock);
498 list_add_tail(&req->queuelist, &nbd->queue_head);
499 spin_unlock(&nbd->queue_lock);
502 nbd->active_req = NULL;
503 mutex_unlock(&nbd->tx_lock);
504 wake_up_all(&nbd->active_wq);
510 nbd_end_request(nbd, req);
513 static int nbd_thread(void *data)
515 struct nbd_device *nbd = data;
518 set_user_nice(current, MIN_NICE);
519 while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
520 /* wait for something to do */
521 wait_event_interruptible(nbd->waiting_wq,
522 kthread_should_stop() ||
523 !list_empty(&nbd->waiting_queue));
525 /* extract request */
526 if (list_empty(&nbd->waiting_queue))
529 spin_lock_irq(&nbd->queue_lock);
530 req = list_entry(nbd->waiting_queue.next, struct request,
532 list_del_init(&req->queuelist);
533 spin_unlock_irq(&nbd->queue_lock);
536 nbd_handle_req(nbd, req);
542 * We always wait for result of write, for now. It would be nice to make it optional
544 * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
545 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
548 static void do_nbd_request(struct request_queue *q)
549 __releases(q->queue_lock) __acquires(q->queue_lock)
553 while ((req = blk_fetch_request(q)) != NULL) {
554 struct nbd_device *nbd;
556 spin_unlock_irq(q->queue_lock);
558 nbd = req->rq_disk->private_data;
560 BUG_ON(nbd->magic != NBD_MAGIC);
562 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
565 if (unlikely(!nbd->sock)) {
566 dev_err(disk_to_dev(nbd->disk),
567 "Attempted send on closed socket\n");
569 nbd_end_request(nbd, req);
570 spin_lock_irq(q->queue_lock);
574 spin_lock_irq(&nbd->queue_lock);
575 list_add_tail(&req->queuelist, &nbd->waiting_queue);
576 spin_unlock_irq(&nbd->queue_lock);
578 wake_up(&nbd->waiting_wq);
580 spin_lock_irq(q->queue_lock);
584 /* Must be called with tx_lock held */
586 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
587 unsigned int cmd, unsigned long arg)
590 case NBD_DISCONNECT: {
593 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
597 mutex_unlock(&nbd->tx_lock);
599 mutex_lock(&nbd->tx_lock);
600 blk_rq_init(NULL, &sreq);
601 sreq.cmd_type = REQ_TYPE_SPECIAL;
602 nbd_cmd(&sreq) = NBD_CMD_DISC;
604 /* Check again after getting mutex back. */
610 nbd_send_req(nbd, &sreq);
614 case NBD_CLEAR_SOCK: {
615 struct socket *sock = nbd->sock;
618 BUG_ON(!list_empty(&nbd->queue_head));
619 BUG_ON(!list_empty(&nbd->waiting_queue));
631 sock = sockfd_lookup(arg, &err);
635 bdev->bd_invalidated = 1;
636 nbd->disconnect = 0; /* we're connected now */
642 case NBD_SET_BLKSIZE:
644 nbd->bytesize &= ~(nbd->blksize-1);
645 bdev->bd_inode->i_size = nbd->bytesize;
646 set_blocksize(bdev, nbd->blksize);
647 set_capacity(nbd->disk, nbd->bytesize >> 9);
651 nbd->bytesize = arg & ~(nbd->blksize-1);
652 bdev->bd_inode->i_size = nbd->bytesize;
653 set_blocksize(bdev, nbd->blksize);
654 set_capacity(nbd->disk, nbd->bytesize >> 9);
657 case NBD_SET_TIMEOUT:
658 nbd->xmit_timeout = arg * HZ;
665 case NBD_SET_SIZE_BLOCKS:
666 nbd->bytesize = ((u64) arg) * nbd->blksize;
667 bdev->bd_inode->i_size = nbd->bytesize;
668 set_blocksize(bdev, nbd->blksize);
669 set_capacity(nbd->disk, nbd->bytesize >> 9);
673 struct task_struct *thread;
682 mutex_unlock(&nbd->tx_lock);
684 if (nbd->flags & NBD_FLAG_READ_ONLY)
685 set_device_ro(bdev, true);
686 if (nbd->flags & NBD_FLAG_SEND_TRIM)
687 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
689 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
690 blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
692 blk_queue_flush(nbd->disk->queue, 0);
694 thread = kthread_run(nbd_thread, nbd, "%s",
695 nbd->disk->disk_name);
696 if (IS_ERR(thread)) {
697 mutex_lock(&nbd->tx_lock);
698 return PTR_ERR(thread);
701 error = nbd_do_it(nbd);
702 kthread_stop(thread);
704 mutex_lock(&nbd->tx_lock);
707 sock_shutdown(nbd, 0);
711 dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
713 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
714 set_device_ro(bdev, false);
719 bdev->bd_inode->i_size = 0;
720 set_capacity(nbd->disk, 0);
722 ioctl_by_bdev(bdev, BLKRRPART, 0);
723 if (nbd->disconnect) /* user requested, ignore socket errors */
725 return nbd->harderror;
730 * This is for compatibility only. The queue is always cleared
731 * by NBD_DO_IT or NBD_CLEAR_SOCK.
735 case NBD_PRINT_DEBUG:
736 dev_info(disk_to_dev(nbd->disk),
737 "next = %p, prev = %p, head = %p\n",
738 nbd->queue_head.next, nbd->queue_head.prev,
745 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
746 unsigned int cmd, unsigned long arg)
748 struct nbd_device *nbd = bdev->bd_disk->private_data;
751 if (!capable(CAP_SYS_ADMIN))
754 BUG_ON(nbd->magic != NBD_MAGIC);
756 mutex_lock(&nbd->tx_lock);
757 error = __nbd_ioctl(bdev, nbd, cmd, arg);
758 mutex_unlock(&nbd->tx_lock);
763 static const struct block_device_operations nbd_fops =
765 .owner = THIS_MODULE,
770 * And here should be modules and kernel interface
771 * (Just smiley confuses emacs :-)
774 static int __init nbd_init(void)
780 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
783 printk(KERN_ERR "nbd: max_part must be >= 0\n");
789 part_shift = fls(max_part);
792 * Adjust max_part according to part_shift as it is exported
793 * to user space so that user can know the max number of
794 * partition kernel should be able to manage.
796 * Note that -1 is required because partition 0 is reserved
797 * for the whole disk.
799 max_part = (1UL << part_shift) - 1;
802 if ((1UL << part_shift) > DISK_MAX_PARTS)
805 if (nbds_max > 1UL << (MINORBITS - part_shift))
808 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
812 for (i = 0; i < nbds_max; i++) {
813 struct gendisk *disk = alloc_disk(1 << part_shift);
816 nbd_dev[i].disk = disk;
818 * The new linux 2.5 block layer implementation requires
819 * every gendisk to have its very own request_queue struct.
820 * These structs are big so we dynamically allocate them.
822 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
828 * Tell the block layer that we are not a rotational device
830 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
831 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
832 disk->queue->limits.discard_granularity = 512;
833 disk->queue->limits.max_discard_sectors = UINT_MAX;
834 disk->queue->limits.discard_zeroes_data = 0;
835 blk_queue_max_hw_sectors(disk->queue, 65536);
836 disk->queue->limits.max_sectors = 256;
839 if (register_blkdev(NBD_MAJOR, "nbd")) {
844 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
846 for (i = 0; i < nbds_max; i++) {
847 struct gendisk *disk = nbd_dev[i].disk;
848 nbd_dev[i].magic = NBD_MAGIC;
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)");