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>
43 #include <linux/nbd.h>
46 #define LO_MAGIC 0x68797548
49 #define dprintk(flags, fmt...)
51 #define dprintk(flags, fmt...) do { \
52 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
54 #define DBG_IOCTL 0x0004
55 #define DBG_INIT 0x0010
56 #define DBG_EXIT 0x0020
57 #define DBG_BLKDEV 0x0100
60 static unsigned int debugflags;
63 static unsigned int nbds_max = 16;
64 static struct nbd_device *nbd_dev;
68 * Use just one lock (or at most 1 per NIC). Two arguments for this:
69 * 1. Each NIC is essentially a synchronization point for all servers
70 * accessed through that NIC so there's no need to have more locks
72 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
73 * down each lock to the point where they're actually slower than just
75 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
77 static DEFINE_SPINLOCK(nbd_lock);
80 static const char *ioctl_cmd_to_ascii(int cmd)
83 case NBD_SET_SOCK: return "set-sock";
84 case NBD_SET_BLKSIZE: return "set-blksize";
85 case NBD_SET_SIZE: return "set-size";
86 case NBD_DO_IT: return "do-it";
87 case NBD_CLEAR_SOCK: return "clear-sock";
88 case NBD_CLEAR_QUE: return "clear-que";
89 case NBD_PRINT_DEBUG: return "print-debug";
90 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
91 case NBD_DISCONNECT: return "disconnect";
92 case NBD_SET_TIMEOUT: return "set-timeout";
93 case NBD_SET_FLAGS: return "set-flags";
94 case BLKROSET: return "set-read-only";
95 case BLKFLSBUF: return "flush-buffer-cache";
100 static const char *nbdcmd_to_ascii(int cmd)
103 case NBD_CMD_READ: return "read";
104 case NBD_CMD_WRITE: return "write";
105 case NBD_CMD_DISC: return "disconnect";
106 case NBD_CMD_FLUSH: return "flush";
112 static void nbd_end_request(struct request *req)
114 int error = req->errors ? -EIO : 0;
115 struct request_queue *q = req->q;
118 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
119 req, error ? "failed" : "done");
121 spin_lock_irqsave(q->queue_lock, flags);
122 __blk_end_request_all(req, error);
123 spin_unlock_irqrestore(q->queue_lock, flags);
126 static void sock_shutdown(struct nbd_device *lo, int lock)
128 /* Forcibly shutdown the socket causing all listeners
131 * FIXME: This code is duplicated from sys_shutdown, but
132 * there should be a more generic interface rather than
133 * calling socket ops directly here */
135 mutex_lock(&lo->tx_lock);
137 printk(KERN_WARNING "%s: shutting down socket\n",
138 lo->disk->disk_name);
139 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
143 mutex_unlock(&lo->tx_lock);
146 static void nbd_xmit_timeout(unsigned long arg)
148 struct task_struct *task = (struct task_struct *)arg;
150 printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
151 task->comm, task->pid);
152 force_sig(SIGKILL, task);
156 * Send or receive packet.
158 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
161 struct socket *sock = lo->sock;
165 sigset_t blocked, oldset;
167 if (unlikely(!sock)) {
168 printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
169 lo->disk->disk_name, (send ? "send" : "recv"));
173 /* Allow interception of SIGKILL only
174 * Don't allow other signals to interrupt the transmission */
175 siginitsetinv(&blocked, sigmask(SIGKILL));
176 sigprocmask(SIG_SETMASK, &blocked, &oldset);
179 sock->sk->sk_allocation = GFP_NOIO;
184 msg.msg_control = NULL;
185 msg.msg_controllen = 0;
186 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
189 struct timer_list ti;
191 if (lo->xmit_timeout) {
193 ti.function = nbd_xmit_timeout;
194 ti.data = (unsigned long)current;
195 ti.expires = jiffies + lo->xmit_timeout;
198 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
199 if (lo->xmit_timeout)
202 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
204 if (signal_pending(current)) {
206 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
207 task_pid_nr(current), current->comm,
208 dequeue_signal_lock(current, ¤t->blocked, &info));
210 sock_shutdown(lo, !send);
216 result = -EPIPE; /* short read */
223 sigprocmask(SIG_SETMASK, &oldset, NULL);
228 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
232 void *kaddr = kmap(bvec->bv_page);
233 result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
234 kunmap(bvec->bv_page);
238 /* always call with the tx_lock held */
239 static int nbd_send_req(struct nbd_device *lo, struct request *req)
242 struct nbd_request request;
243 unsigned long size = blk_rq_bytes(req);
245 request.magic = htonl(NBD_REQUEST_MAGIC);
246 /* If FUA is set in the request, and we are told to send FUA, then OR in NBD_CMD_FLAG_FUA */
247 request.type = htonl(nbd_cmd(req) |
248 (( (req->cmd_flags & REQ_FUA) && (lo->flags & NBD_FLAG_SEND_FUA)) ?
249 NBD_CMD_FLAG_FUA : 0));
250 /* Send from & len as zero on FLUSH - other values reserved per protocol */
251 if (request.type == NBD_CMD_FLUSH) {
255 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
256 request.len = htonl(size);
258 memcpy(request.handle, &req, sizeof(req));
260 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
261 lo->disk->disk_name, req,
262 nbdcmd_to_ascii(nbd_cmd(req)),
263 (unsigned long long)blk_rq_pos(req) << 9,
265 result = sock_xmit(lo, 1, &request, sizeof(request),
266 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
268 printk(KERN_ERR "%s: Send control failed (result %d)\n",
269 lo->disk->disk_name, result);
273 if (nbd_cmd(req) == NBD_CMD_WRITE) {
274 struct req_iterator iter;
275 struct bio_vec *bvec;
277 * we are really probing at internals to determine
278 * whether to set MSG_MORE or not...
280 rq_for_each_segment(bvec, req, iter) {
282 if (!rq_iter_last(req, iter))
284 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
285 lo->disk->disk_name, req, bvec->bv_len);
286 result = sock_send_bvec(lo, bvec, flags);
288 printk(KERN_ERR "%s: Send data failed (result %d)\n",
289 lo->disk->disk_name, result);
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 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
347 lo->disk->disk_name, result);
351 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
352 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
354 (unsigned long)ntohl(reply.magic));
359 req = nbd_find_request(lo, *(struct request **)reply.handle);
361 result = PTR_ERR(req);
362 if (result != -ENOENT)
365 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
366 lo->disk->disk_name, reply.handle);
371 if (ntohl(reply.error)) {
372 printk(KERN_ERR "%s: Other side returned error (%d)\n",
373 lo->disk->disk_name, ntohl(reply.error));
378 dprintk(DBG_RX, "%s: request %p: got reply\n",
379 lo->disk->disk_name, req);
380 if (nbd_cmd(req) == NBD_CMD_READ) {
381 struct req_iterator iter;
382 struct bio_vec *bvec;
384 rq_for_each_segment(bvec, req, iter) {
385 result = sock_recv_bvec(lo, bvec);
387 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
388 lo->disk->disk_name, result);
392 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
393 lo->disk->disk_name, req, bvec->bv_len);
398 lo->harderror = result;
402 static ssize_t pid_show(struct device *dev,
403 struct device_attribute *attr, char *buf)
405 struct gendisk *disk = dev_to_disk(dev);
407 return sprintf(buf, "%ld\n",
408 (long) ((struct nbd_device *)disk->private_data)->pid);
411 static struct device_attribute pid_attr = {
412 .attr = { .name = "pid", .mode = S_IRUGO},
416 static int nbd_do_it(struct nbd_device *lo)
421 BUG_ON(lo->magic != LO_MAGIC);
423 lo->pid = current->pid;
424 ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
426 printk(KERN_ERR "nbd: sysfs_create_file failed!");
431 while ((req = nbd_read_stat(lo)) != NULL)
432 nbd_end_request(req);
434 sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
439 static void nbd_clear_que(struct nbd_device *lo)
443 BUG_ON(lo->magic != LO_MAGIC);
446 * Because we have set lo->sock to NULL under the tx_lock, all
447 * modifications to the list must have completed by now. For
448 * the same reason, the active_req must be NULL.
450 * As a consequence, we don't need to take the spin lock while
451 * purging the list here.
454 BUG_ON(lo->active_req);
456 while (!list_empty(&lo->queue_head)) {
457 req = list_entry(lo->queue_head.next, struct request,
459 list_del_init(&req->queuelist);
461 nbd_end_request(req);
466 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
468 if (req->cmd_type != REQ_TYPE_FS)
471 nbd_cmd(req) = NBD_CMD_READ;
472 if (rq_data_dir(req) == WRITE) {
473 nbd_cmd(req) = NBD_CMD_WRITE;
474 if (lo->flags & NBD_FLAG_READ_ONLY) {
475 printk(KERN_ERR "%s: Write on read-only\n",
476 lo->disk->disk_name);
481 if (req->cmd_flags & REQ_FLUSH) {
482 if (unlikely(blk_rq_sectors(req))) {
483 /* Elevator is meant to guarantee that a request with REQ_FLUSH
484 * set is broken into an empty request with REQ_FLUSH set then
485 * the rest of the content (if any). If this doesn't happen,
486 * whinge, then proceed to do the content without a flush.
488 printk(KERN_ERR "%s: nbd passed non-empty flush request\n",
489 lo->disk->disk_name);
492 if (lo->flags & NBD_FLAG_SEND_FLUSH)
493 nbd_cmd(req) = NBD_CMD_FLUSH;
495 /* Ignore flush that we don't need */
496 nbd_end_request(req);
504 mutex_lock(&lo->tx_lock);
505 if (unlikely(!lo->sock)) {
506 mutex_unlock(&lo->tx_lock);
507 printk(KERN_ERR "%s: Attempted send on closed socket\n",
508 lo->disk->disk_name);
512 lo->active_req = req;
514 if (nbd_send_req(lo, req) != 0) {
515 printk(KERN_ERR "%s: Request send failed\n",
516 lo->disk->disk_name);
518 nbd_end_request(req);
520 spin_lock(&lo->queue_lock);
521 list_add(&req->queuelist, &lo->queue_head);
522 spin_unlock(&lo->queue_lock);
525 lo->active_req = NULL;
526 mutex_unlock(&lo->tx_lock);
527 wake_up_all(&lo->active_wq);
533 nbd_end_request(req);
536 static int nbd_thread(void *data)
538 struct nbd_device *lo = data;
541 set_user_nice(current, -20);
542 while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
543 /* wait for something to do */
544 wait_event_interruptible(lo->waiting_wq,
545 kthread_should_stop() ||
546 !list_empty(&lo->waiting_queue));
548 /* extract request */
549 if (list_empty(&lo->waiting_queue))
552 spin_lock_irq(&lo->queue_lock);
553 req = list_entry(lo->waiting_queue.next, struct request,
555 list_del_init(&req->queuelist);
556 spin_unlock_irq(&lo->queue_lock);
559 nbd_handle_req(lo, req);
565 * We always wait for result of write, for now. It would be nice to make it optional
567 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
568 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
571 static void do_nbd_request(struct request_queue *q)
575 while ((req = blk_fetch_request(q)) != NULL) {
576 struct nbd_device *lo;
578 spin_unlock_irq(q->queue_lock);
580 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
581 req->rq_disk->disk_name, req, req->cmd_type);
583 lo = req->rq_disk->private_data;
585 BUG_ON(lo->magic != LO_MAGIC);
587 if (unlikely(!lo->sock)) {
588 printk(KERN_ERR "%s: Attempted send on closed socket\n",
589 lo->disk->disk_name);
591 nbd_end_request(req);
592 spin_lock_irq(q->queue_lock);
596 spin_lock_irq(&lo->queue_lock);
597 list_add_tail(&req->queuelist, &lo->waiting_queue);
598 spin_unlock_irq(&lo->queue_lock);
600 wake_up(&lo->waiting_wq);
602 spin_lock_irq(q->queue_lock);
606 /* Must be called with tx_lock held */
608 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
609 unsigned int cmd, unsigned long arg)
612 case NBD_DISCONNECT: {
615 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
617 blk_rq_init(NULL, &sreq);
618 sreq.cmd_type = REQ_TYPE_SPECIAL;
619 nbd_cmd(&sreq) = NBD_CMD_DISC;
622 nbd_send_req(lo, &sreq);
626 case NBD_CLEAR_SOCK: {
633 BUG_ON(!list_empty(&lo->queue_head));
645 struct inode *inode = file->f_path.dentry->d_inode;
646 if (S_ISSOCK(inode->i_mode)) {
648 lo->sock = SOCKET_I(inode);
650 bdev->bd_invalidated = 1;
659 case NBD_SET_BLKSIZE:
661 lo->bytesize &= ~(lo->blksize-1);
662 bdev->bd_inode->i_size = lo->bytesize;
663 set_blocksize(bdev, lo->blksize);
664 set_capacity(lo->disk, lo->bytesize >> 9);
668 lo->bytesize = arg & ~(lo->blksize-1);
669 bdev->bd_inode->i_size = lo->bytesize;
670 set_blocksize(bdev, lo->blksize);
671 set_capacity(lo->disk, lo->bytesize >> 9);
674 case NBD_SET_TIMEOUT:
675 lo->xmit_timeout = arg * HZ;
682 if (lo->flags & NBD_FLAG_ROTATIONAL)
683 queue_flag_clear_unlocked(QUEUE_FLAG_NONROT, lo->disk->queue);
685 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, lo->disk->queue);
690 case NBD_SET_SIZE_BLOCKS:
691 lo->bytesize = ((u64) arg) * lo->blksize;
692 bdev->bd_inode->i_size = lo->bytesize;
693 set_blocksize(bdev, lo->blksize);
694 set_capacity(lo->disk, lo->bytesize >> 9);
698 struct task_struct *thread;
707 mutex_unlock(&lo->tx_lock);
709 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
710 if (IS_ERR(thread)) {
711 mutex_lock(&lo->tx_lock);
712 return PTR_ERR(thread);
714 wake_up_process(thread);
715 error = nbd_do_it(lo);
716 kthread_stop(thread);
718 mutex_lock(&lo->tx_lock);
721 sock_shutdown(lo, 0);
725 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
729 bdev->bd_inode->i_size = 0;
730 set_capacity(lo->disk, 0);
732 ioctl_by_bdev(bdev, BLKRRPART, 0);
733 return lo->harderror;
738 * This is for compatibility only. The queue is always cleared
739 * by NBD_DO_IT or NBD_CLEAR_SOCK.
741 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
744 case NBD_PRINT_DEBUG:
745 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
746 bdev->bd_disk->disk_name,
747 lo->queue_head.next, lo->queue_head.prev,
754 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
755 unsigned int cmd, unsigned long arg)
757 struct nbd_device *lo = bdev->bd_disk->private_data;
760 if (!capable(CAP_SYS_ADMIN))
763 BUG_ON(lo->magic != LO_MAGIC);
765 /* Anyone capable of this syscall can do *real bad* things */
766 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
767 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
769 mutex_lock(&lo->tx_lock);
770 error = __nbd_ioctl(bdev, lo, cmd, arg);
771 mutex_unlock(&lo->tx_lock);
776 static const struct block_device_operations nbd_fops =
778 .owner = THIS_MODULE,
783 * And here should be modules and kernel interface
784 * (Just smiley confuses emacs :-)
787 static int __init nbd_init(void)
793 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
796 printk(KERN_CRIT "nbd: max_part must be >= 0\n");
800 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
806 part_shift = fls(max_part);
808 for (i = 0; i < nbds_max; i++) {
809 struct gendisk *disk = alloc_disk(1 << part_shift);
812 nbd_dev[i].disk = disk;
814 * The new linux 2.5 block layer implementation requires
815 * every gendisk to have its very own request_queue struct.
816 * These structs are big so we dynamically allocate them.
818 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
823 /* In order not to confuse the elevator, say we always
824 * want FLUSH and FUA. We won't send them to the server
825 * unless the relevant flag bit is set
827 blk_queue_flush(disk->queue, REQ_FLUSH | REQ_FUA);
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");