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@suse.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/compiler.h>
28 #include <linux/err.h>
29 #include <linux/kernel.h>
30 #include <linux/slab.h>
32 #include <linux/net.h>
33 #include <linux/kthread.h>
35 #include <asm/uaccess.h>
36 #include <asm/system.h>
37 #include <asm/types.h>
39 #include <linux/nbd.h>
41 #define LO_MAGIC 0x68797548
44 #define dprintk(flags, fmt...)
46 #define dprintk(flags, fmt...) do { \
47 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
49 #define DBG_IOCTL 0x0004
50 #define DBG_INIT 0x0010
51 #define DBG_EXIT 0x0020
52 #define DBG_BLKDEV 0x0100
55 static unsigned int debugflags;
58 static unsigned int nbds_max = 16;
59 static struct nbd_device *nbd_dev;
63 * Use just one lock (or at most 1 per NIC). Two arguments for this:
64 * 1. Each NIC is essentially a synchronization point for all servers
65 * accessed through that NIC so there's no need to have more locks
67 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
68 * down each lock to the point where they're actually slower than just
70 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
72 static DEFINE_SPINLOCK(nbd_lock);
75 static const char *ioctl_cmd_to_ascii(int cmd)
78 case NBD_SET_SOCK: return "set-sock";
79 case NBD_SET_BLKSIZE: return "set-blksize";
80 case NBD_SET_SIZE: return "set-size";
81 case NBD_DO_IT: return "do-it";
82 case NBD_CLEAR_SOCK: return "clear-sock";
83 case NBD_CLEAR_QUE: return "clear-que";
84 case NBD_PRINT_DEBUG: return "print-debug";
85 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
86 case NBD_DISCONNECT: return "disconnect";
87 case BLKROSET: return "set-read-only";
88 case BLKFLSBUF: return "flush-buffer-cache";
93 static const char *nbdcmd_to_ascii(int cmd)
96 case NBD_CMD_READ: return "read";
97 case NBD_CMD_WRITE: return "write";
98 case NBD_CMD_DISC: return "disconnect";
104 static void nbd_end_request(struct request *req)
106 int error = req->errors ? -EIO : 0;
107 struct request_queue *q = req->q;
110 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
111 req, error ? "failed" : "done");
113 spin_lock_irqsave(q->queue_lock, flags);
114 __blk_end_request_all(req, error);
115 spin_unlock_irqrestore(q->queue_lock, flags);
118 static void sock_shutdown(struct nbd_device *lo, int lock)
120 /* Forcibly shutdown the socket causing all listeners
123 * FIXME: This code is duplicated from sys_shutdown, but
124 * there should be a more generic interface rather than
125 * calling socket ops directly here */
127 mutex_lock(&lo->tx_lock);
129 printk(KERN_WARNING "%s: shutting down socket\n",
130 lo->disk->disk_name);
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 printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
161 lo->disk->disk_name, (send ? "send" : "recv"));
165 /* Allow interception of SIGKILL only
166 * Don't allow other signals to interrupt the transmission */
167 siginitsetinv(&blocked, sigmask(SIGKILL));
168 sigprocmask(SIG_SETMASK, &blocked, &oldset);
171 sock->sk->sk_allocation = GFP_NOIO;
176 msg.msg_control = NULL;
177 msg.msg_controllen = 0;
178 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
181 struct timer_list ti;
183 if (lo->xmit_timeout) {
185 ti.function = nbd_xmit_timeout;
186 ti.data = (unsigned long)current;
187 ti.expires = jiffies + lo->xmit_timeout;
190 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
191 if (lo->xmit_timeout)
194 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
196 if (signal_pending(current)) {
198 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
199 task_pid_nr(current), current->comm,
200 dequeue_signal_lock(current, ¤t->blocked, &info));
202 sock_shutdown(lo, !send);
208 result = -EPIPE; /* short read */
215 sigprocmask(SIG_SETMASK, &oldset, NULL);
220 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
224 void *kaddr = kmap(bvec->bv_page);
225 result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
226 kunmap(bvec->bv_page);
230 /* always call with the tx_lock held */
231 static int nbd_send_req(struct nbd_device *lo, struct request *req)
234 struct nbd_request request;
235 unsigned long size = blk_rq_bytes(req);
237 request.magic = htonl(NBD_REQUEST_MAGIC);
238 request.type = htonl(nbd_cmd(req));
239 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
240 request.len = htonl(size);
241 memcpy(request.handle, &req, sizeof(req));
243 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
244 lo->disk->disk_name, req,
245 nbdcmd_to_ascii(nbd_cmd(req)),
246 (unsigned long long)blk_rq_pos(req) << 9,
248 result = sock_xmit(lo, 1, &request, sizeof(request),
249 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
251 printk(KERN_ERR "%s: Send control failed (result %d)\n",
252 lo->disk->disk_name, result);
256 if (nbd_cmd(req) == NBD_CMD_WRITE) {
257 struct req_iterator iter;
258 struct bio_vec *bvec;
260 * we are really probing at internals to determine
261 * whether to set MSG_MORE or not...
263 rq_for_each_segment(bvec, req, iter) {
265 if (!rq_iter_last(req, iter))
267 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
268 lo->disk->disk_name, req, bvec->bv_len);
269 result = sock_send_bvec(lo, bvec, flags);
271 printk(KERN_ERR "%s: Send data failed (result %d)\n",
272 lo->disk->disk_name, result);
283 static struct request *nbd_find_request(struct nbd_device *lo,
284 struct request *xreq)
286 struct request *req, *tmp;
289 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
293 spin_lock(&lo->queue_lock);
294 list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
297 list_del_init(&req->queuelist);
298 spin_unlock(&lo->queue_lock);
301 spin_unlock(&lo->queue_lock);
309 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
312 void *kaddr = kmap(bvec->bv_page);
313 result = sock_xmit(lo, 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 *lo)
323 struct nbd_reply reply;
327 result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
329 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
330 lo->disk->disk_name, result);
334 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
335 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
337 (unsigned long)ntohl(reply.magic));
342 req = nbd_find_request(lo, *(struct request **)reply.handle);
344 result = PTR_ERR(req);
345 if (result != -ENOENT)
348 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
349 lo->disk->disk_name, reply.handle);
354 if (ntohl(reply.error)) {
355 printk(KERN_ERR "%s: Other side returned error (%d)\n",
356 lo->disk->disk_name, ntohl(reply.error));
361 dprintk(DBG_RX, "%s: request %p: got reply\n",
362 lo->disk->disk_name, req);
363 if (nbd_cmd(req) == NBD_CMD_READ) {
364 struct req_iterator iter;
365 struct bio_vec *bvec;
367 rq_for_each_segment(bvec, req, iter) {
368 result = sock_recv_bvec(lo, bvec);
370 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
371 lo->disk->disk_name, result);
375 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
376 lo->disk->disk_name, req, bvec->bv_len);
381 lo->harderror = result;
385 static ssize_t pid_show(struct device *dev,
386 struct device_attribute *attr, char *buf)
388 struct gendisk *disk = dev_to_disk(dev);
390 return sprintf(buf, "%ld\n",
391 (long) ((struct nbd_device *)disk->private_data)->pid);
394 static struct device_attribute pid_attr = {
395 .attr = { .name = "pid", .mode = S_IRUGO},
399 static int nbd_do_it(struct nbd_device *lo)
404 BUG_ON(lo->magic != LO_MAGIC);
406 lo->pid = current->pid;
407 ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
409 printk(KERN_ERR "nbd: sysfs_create_file failed!");
414 while ((req = nbd_read_stat(lo)) != NULL)
415 nbd_end_request(req);
417 sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
422 static void nbd_clear_que(struct nbd_device *lo)
426 BUG_ON(lo->magic != LO_MAGIC);
429 * Because we have set lo->sock to NULL under the tx_lock, all
430 * modifications to the list must have completed by now. For
431 * the same reason, the active_req must be NULL.
433 * As a consequence, we don't need to take the spin lock while
434 * purging the list here.
437 BUG_ON(lo->active_req);
439 while (!list_empty(&lo->queue_head)) {
440 req = list_entry(lo->queue_head.next, struct request,
442 list_del_init(&req->queuelist);
444 nbd_end_request(req);
449 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
451 if (!blk_fs_request(req))
454 nbd_cmd(req) = NBD_CMD_READ;
455 if (rq_data_dir(req) == WRITE) {
456 nbd_cmd(req) = NBD_CMD_WRITE;
457 if (lo->flags & NBD_READ_ONLY) {
458 printk(KERN_ERR "%s: Write on read-only\n",
459 lo->disk->disk_name);
466 mutex_lock(&lo->tx_lock);
467 if (unlikely(!lo->sock)) {
468 mutex_unlock(&lo->tx_lock);
469 printk(KERN_ERR "%s: Attempted send on closed socket\n",
470 lo->disk->disk_name);
474 lo->active_req = req;
476 if (nbd_send_req(lo, req) != 0) {
477 printk(KERN_ERR "%s: Request send failed\n",
478 lo->disk->disk_name);
480 nbd_end_request(req);
482 spin_lock(&lo->queue_lock);
483 list_add(&req->queuelist, &lo->queue_head);
484 spin_unlock(&lo->queue_lock);
487 lo->active_req = NULL;
488 mutex_unlock(&lo->tx_lock);
489 wake_up_all(&lo->active_wq);
495 nbd_end_request(req);
498 static int nbd_thread(void *data)
500 struct nbd_device *lo = data;
503 set_user_nice(current, -20);
504 while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
505 /* wait for something to do */
506 wait_event_interruptible(lo->waiting_wq,
507 kthread_should_stop() ||
508 !list_empty(&lo->waiting_queue));
510 /* extract request */
511 if (list_empty(&lo->waiting_queue))
514 spin_lock_irq(&lo->queue_lock);
515 req = list_entry(lo->waiting_queue.next, struct request,
517 list_del_init(&req->queuelist);
518 spin_unlock_irq(&lo->queue_lock);
521 nbd_handle_req(lo, req);
527 * We always wait for result of write, for now. It would be nice to make it optional
529 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
530 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
533 static void do_nbd_request(struct request_queue *q)
537 while ((req = blk_fetch_request(q)) != NULL) {
538 struct nbd_device *lo;
540 spin_unlock_irq(q->queue_lock);
542 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
543 req->rq_disk->disk_name, req, req->cmd_type);
545 lo = req->rq_disk->private_data;
547 BUG_ON(lo->magic != LO_MAGIC);
549 if (unlikely(!lo->sock)) {
550 printk(KERN_ERR "%s: Attempted send on closed socket\n",
551 lo->disk->disk_name);
553 nbd_end_request(req);
554 spin_lock_irq(q->queue_lock);
558 spin_lock_irq(&lo->queue_lock);
559 list_add_tail(&req->queuelist, &lo->waiting_queue);
560 spin_unlock_irq(&lo->queue_lock);
562 wake_up(&lo->waiting_wq);
564 spin_lock_irq(q->queue_lock);
568 /* Must be called with tx_lock held */
570 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
571 unsigned int cmd, unsigned long arg)
574 case NBD_DISCONNECT: {
577 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
579 blk_rq_init(NULL, &sreq);
580 sreq.cmd_type = REQ_TYPE_SPECIAL;
581 nbd_cmd(&sreq) = NBD_CMD_DISC;
584 nbd_send_req(lo, &sreq);
588 case NBD_CLEAR_SOCK: {
595 BUG_ON(!list_empty(&lo->queue_head));
607 struct inode *inode = file->f_path.dentry->d_inode;
608 if (S_ISSOCK(inode->i_mode)) {
610 lo->sock = SOCKET_I(inode);
612 bdev->bd_invalidated = 1;
621 case NBD_SET_BLKSIZE:
623 lo->bytesize &= ~(lo->blksize-1);
624 bdev->bd_inode->i_size = lo->bytesize;
625 set_blocksize(bdev, lo->blksize);
626 set_capacity(lo->disk, lo->bytesize >> 9);
630 lo->bytesize = arg & ~(lo->blksize-1);
631 bdev->bd_inode->i_size = lo->bytesize;
632 set_blocksize(bdev, lo->blksize);
633 set_capacity(lo->disk, lo->bytesize >> 9);
636 case NBD_SET_TIMEOUT:
637 lo->xmit_timeout = arg * HZ;
640 case NBD_SET_SIZE_BLOCKS:
641 lo->bytesize = ((u64) arg) * lo->blksize;
642 bdev->bd_inode->i_size = lo->bytesize;
643 set_blocksize(bdev, lo->blksize);
644 set_capacity(lo->disk, lo->bytesize >> 9);
648 struct task_struct *thread;
657 mutex_unlock(&lo->tx_lock);
659 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
660 if (IS_ERR(thread)) {
661 mutex_lock(&lo->tx_lock);
662 return PTR_ERR(thread);
664 wake_up_process(thread);
665 error = nbd_do_it(lo);
666 kthread_stop(thread);
668 mutex_lock(&lo->tx_lock);
671 sock_shutdown(lo, 0);
675 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
679 bdev->bd_inode->i_size = 0;
680 set_capacity(lo->disk, 0);
682 ioctl_by_bdev(bdev, BLKRRPART, 0);
683 return lo->harderror;
688 * This is for compatibility only. The queue is always cleared
689 * by NBD_DO_IT or NBD_CLEAR_SOCK.
691 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
694 case NBD_PRINT_DEBUG:
695 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
696 bdev->bd_disk->disk_name,
697 lo->queue_head.next, lo->queue_head.prev,
704 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
705 unsigned int cmd, unsigned long arg)
707 struct nbd_device *lo = bdev->bd_disk->private_data;
710 if (!capable(CAP_SYS_ADMIN))
713 BUG_ON(lo->magic != LO_MAGIC);
715 /* Anyone capable of this syscall can do *real bad* things */
716 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
717 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
719 mutex_lock(&lo->tx_lock);
720 error = __nbd_ioctl(bdev, lo, cmd, arg);
721 mutex_unlock(&lo->tx_lock);
726 static const struct block_device_operations nbd_fops =
728 .owner = THIS_MODULE,
729 .locked_ioctl = nbd_ioctl,
733 * And here should be modules and kernel interface
734 * (Just smiley confuses emacs :-)
737 static int __init nbd_init(void)
743 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
746 printk(KERN_CRIT "nbd: max_part must be >= 0\n");
750 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
756 part_shift = fls(max_part);
758 for (i = 0; i < nbds_max; i++) {
759 struct gendisk *disk = alloc_disk(1 << part_shift);
762 nbd_dev[i].disk = disk;
764 * The new linux 2.5 block layer implementation requires
765 * every gendisk to have its very own request_queue struct.
766 * These structs are big so we dynamically allocate them.
768 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
774 * Tell the block layer that we are not a rotational device
776 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
779 if (register_blkdev(NBD_MAJOR, "nbd")) {
784 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
785 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
787 for (i = 0; i < nbds_max; i++) {
788 struct gendisk *disk = nbd_dev[i].disk;
789 nbd_dev[i].file = NULL;
790 nbd_dev[i].magic = LO_MAGIC;
791 nbd_dev[i].flags = 0;
792 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
793 spin_lock_init(&nbd_dev[i].queue_lock);
794 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
795 mutex_init(&nbd_dev[i].tx_lock);
796 init_waitqueue_head(&nbd_dev[i].active_wq);
797 init_waitqueue_head(&nbd_dev[i].waiting_wq);
798 nbd_dev[i].blksize = 1024;
799 nbd_dev[i].bytesize = 0;
800 disk->major = NBD_MAJOR;
801 disk->first_minor = i << part_shift;
802 disk->fops = &nbd_fops;
803 disk->private_data = &nbd_dev[i];
804 sprintf(disk->disk_name, "nbd%d", i);
805 set_capacity(disk, 0);
812 blk_cleanup_queue(nbd_dev[i].disk->queue);
813 put_disk(nbd_dev[i].disk);
819 static void __exit nbd_cleanup(void)
822 for (i = 0; i < nbds_max; i++) {
823 struct gendisk *disk = nbd_dev[i].disk;
824 nbd_dev[i].magic = 0;
827 blk_cleanup_queue(disk->queue);
831 unregister_blkdev(NBD_MAJOR, "nbd");
833 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
836 module_init(nbd_init);
837 module_exit(nbd_cleanup);
839 MODULE_DESCRIPTION("Network Block Device");
840 MODULE_LICENSE("GPL");
842 module_param(nbds_max, int, 0444);
843 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
844 module_param(max_part, int, 0444);
845 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
847 module_param(debugflags, int, 0644);
848 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");