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 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/compiler.h>
28 #include <linux/err.h>
29 #include <linux/kernel.h>
31 #include <linux/net.h>
32 #include <linux/kthread.h>
34 #include <asm/uaccess.h>
35 #include <asm/system.h>
36 #include <asm/types.h>
38 #include <linux/nbd.h>
40 #define LO_MAGIC 0x68797548
43 #define dprintk(flags, fmt...)
45 #define dprintk(flags, fmt...) do { \
46 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
48 #define DBG_IOCTL 0x0004
49 #define DBG_INIT 0x0010
50 #define DBG_EXIT 0x0020
51 #define DBG_BLKDEV 0x0100
54 static unsigned int debugflags;
57 static unsigned int nbds_max = 16;
58 static struct nbd_device *nbd_dev;
61 * Use just one lock (or at most 1 per NIC). Two arguments for this:
62 * 1. Each NIC is essentially a synchronization point for all servers
63 * accessed through that NIC so there's no need to have more locks
65 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
66 * down each lock to the point where they're actually slower than just
68 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
70 static DEFINE_SPINLOCK(nbd_lock);
73 static const char *ioctl_cmd_to_ascii(int cmd)
76 case NBD_SET_SOCK: return "set-sock";
77 case NBD_SET_BLKSIZE: return "set-blksize";
78 case NBD_SET_SIZE: return "set-size";
79 case NBD_DO_IT: return "do-it";
80 case NBD_CLEAR_SOCK: return "clear-sock";
81 case NBD_CLEAR_QUE: return "clear-que";
82 case NBD_PRINT_DEBUG: return "print-debug";
83 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
84 case NBD_DISCONNECT: return "disconnect";
85 case BLKROSET: return "set-read-only";
86 case BLKFLSBUF: return "flush-buffer-cache";
91 static const char *nbdcmd_to_ascii(int cmd)
94 case NBD_CMD_READ: return "read";
95 case NBD_CMD_WRITE: return "write";
96 case NBD_CMD_DISC: return "disconnect";
102 static void nbd_end_request(struct request *req)
104 int error = req->errors ? -EIO : 0;
105 struct request_queue *q = req->q;
108 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
109 req, error ? "failed" : "done");
111 spin_lock_irqsave(q->queue_lock, flags);
112 __blk_end_request(req, error, req->nr_sectors << 9);
113 spin_unlock_irqrestore(q->queue_lock, flags);
116 static void sock_shutdown(struct nbd_device *lo, int lock)
118 /* Forcibly shutdown the socket causing all listeners
121 * FIXME: This code is duplicated from sys_shutdown, but
122 * there should be a more generic interface rather than
123 * calling socket ops directly here */
125 mutex_lock(&lo->tx_lock);
127 printk(KERN_WARNING "%s: shutting down socket\n",
128 lo->disk->disk_name);
129 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
133 mutex_unlock(&lo->tx_lock);
136 static void nbd_xmit_timeout(unsigned long arg)
138 struct task_struct *task = (struct task_struct *)arg;
140 printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
141 task->comm, task->pid);
142 force_sig(SIGKILL, task);
146 * Send or receive packet.
148 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
151 struct socket *sock = lo->sock;
155 sigset_t blocked, oldset;
157 if (unlikely(!sock)) {
158 printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
159 lo->disk->disk_name, (send ? "send" : "recv"));
163 /* Allow interception of SIGKILL only
164 * Don't allow other signals to interrupt the transmission */
165 siginitsetinv(&blocked, sigmask(SIGKILL));
166 sigprocmask(SIG_SETMASK, &blocked, &oldset);
169 sock->sk->sk_allocation = GFP_NOIO;
174 msg.msg_control = NULL;
175 msg.msg_controllen = 0;
176 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
179 struct timer_list ti;
181 if (lo->xmit_timeout) {
183 ti.function = nbd_xmit_timeout;
184 ti.data = (unsigned long)current;
185 ti.expires = jiffies + lo->xmit_timeout;
188 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
189 if (lo->xmit_timeout)
192 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
194 if (signal_pending(current)) {
196 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
197 task_pid_nr(current), current->comm,
198 dequeue_signal_lock(current, ¤t->blocked, &info));
200 sock_shutdown(lo, !send);
206 result = -EPIPE; /* short read */
213 sigprocmask(SIG_SETMASK, &oldset, NULL);
218 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
222 void *kaddr = kmap(bvec->bv_page);
223 result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
224 kunmap(bvec->bv_page);
228 /* always call with the tx_lock held */
229 static int nbd_send_req(struct nbd_device *lo, struct request *req)
232 struct nbd_request request;
233 unsigned long size = req->nr_sectors << 9;
235 request.magic = htonl(NBD_REQUEST_MAGIC);
236 request.type = htonl(nbd_cmd(req));
237 request.from = cpu_to_be64((u64) req->sector << 9);
238 request.len = htonl(size);
239 memcpy(request.handle, &req, sizeof(req));
241 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
242 lo->disk->disk_name, req,
243 nbdcmd_to_ascii(nbd_cmd(req)),
244 (unsigned long long)req->sector << 9,
245 req->nr_sectors << 9);
246 result = sock_xmit(lo, 1, &request, sizeof(request),
247 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
249 printk(KERN_ERR "%s: Send control failed (result %d)\n",
250 lo->disk->disk_name, result);
254 if (nbd_cmd(req) == NBD_CMD_WRITE) {
255 struct req_iterator iter;
256 struct bio_vec *bvec;
258 * we are really probing at internals to determine
259 * whether to set MSG_MORE or not...
261 rq_for_each_segment(bvec, req, iter) {
263 if (!rq_iter_last(req, iter))
265 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
266 lo->disk->disk_name, req, bvec->bv_len);
267 result = sock_send_bvec(lo, bvec, flags);
269 printk(KERN_ERR "%s: Send data failed (result %d)\n",
270 lo->disk->disk_name, result);
281 static struct request *nbd_find_request(struct nbd_device *lo,
282 struct request *xreq)
284 struct request *req, *tmp;
287 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
291 spin_lock(&lo->queue_lock);
292 list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
295 list_del_init(&req->queuelist);
296 spin_unlock(&lo->queue_lock);
299 spin_unlock(&lo->queue_lock);
307 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
310 void *kaddr = kmap(bvec->bv_page);
311 result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
313 kunmap(bvec->bv_page);
317 /* NULL returned = something went wrong, inform userspace */
318 static struct request *nbd_read_stat(struct nbd_device *lo)
321 struct nbd_reply reply;
325 result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
327 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
328 lo->disk->disk_name, result);
332 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
333 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
335 (unsigned long)ntohl(reply.magic));
340 req = nbd_find_request(lo, *(struct request **)reply.handle);
341 if (unlikely(IS_ERR(req))) {
342 result = PTR_ERR(req);
343 if (result != -ENOENT)
346 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
347 lo->disk->disk_name, reply.handle);
352 if (ntohl(reply.error)) {
353 printk(KERN_ERR "%s: Other side returned error (%d)\n",
354 lo->disk->disk_name, ntohl(reply.error));
359 dprintk(DBG_RX, "%s: request %p: got reply\n",
360 lo->disk->disk_name, req);
361 if (nbd_cmd(req) == NBD_CMD_READ) {
362 struct req_iterator iter;
363 struct bio_vec *bvec;
365 rq_for_each_segment(bvec, req, iter) {
366 result = sock_recv_bvec(lo, bvec);
368 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
369 lo->disk->disk_name, result);
373 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
374 lo->disk->disk_name, req, bvec->bv_len);
379 lo->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, .owner = THIS_MODULE },
397 static int nbd_do_it(struct nbd_device *lo)
402 BUG_ON(lo->magic != LO_MAGIC);
404 lo->pid = current->pid;
405 ret = sysfs_create_file(&lo->disk->dev.kobj, &pid_attr.attr);
407 printk(KERN_ERR "nbd: sysfs_create_file failed!");
411 while ((req = nbd_read_stat(lo)) != NULL)
412 nbd_end_request(req);
414 sysfs_remove_file(&lo->disk->dev.kobj, &pid_attr.attr);
418 static void nbd_clear_que(struct nbd_device *lo)
422 BUG_ON(lo->magic != LO_MAGIC);
425 * Because we have set lo->sock to NULL under the tx_lock, all
426 * modifications to the list must have completed by now. For
427 * the same reason, the active_req must be NULL.
429 * As a consequence, we don't need to take the spin lock while
430 * purging the list here.
433 BUG_ON(lo->active_req);
435 while (!list_empty(&lo->queue_head)) {
436 req = list_entry(lo->queue_head.next, struct request,
438 list_del_init(&req->queuelist);
440 nbd_end_request(req);
445 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
447 if (!blk_fs_request(req))
450 nbd_cmd(req) = NBD_CMD_READ;
451 if (rq_data_dir(req) == WRITE) {
452 nbd_cmd(req) = NBD_CMD_WRITE;
453 if (lo->flags & NBD_READ_ONLY) {
454 printk(KERN_ERR "%s: Write on read-only\n",
455 lo->disk->disk_name);
462 mutex_lock(&lo->tx_lock);
463 if (unlikely(!lo->sock)) {
464 mutex_unlock(&lo->tx_lock);
465 printk(KERN_ERR "%s: Attempted send on closed socket\n",
466 lo->disk->disk_name);
468 nbd_end_request(req);
472 lo->active_req = req;
474 if (nbd_send_req(lo, req) != 0) {
475 printk(KERN_ERR "%s: Request send failed\n",
476 lo->disk->disk_name);
478 nbd_end_request(req);
480 spin_lock(&lo->queue_lock);
481 list_add(&req->queuelist, &lo->queue_head);
482 spin_unlock(&lo->queue_lock);
485 lo->active_req = NULL;
486 mutex_unlock(&lo->tx_lock);
487 wake_up_all(&lo->active_wq);
493 nbd_end_request(req);
496 static int nbd_thread(void *data)
498 struct nbd_device *lo = data;
501 set_user_nice(current, -20);
502 while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
503 /* wait for something to do */
504 wait_event_interruptible(lo->waiting_wq,
505 kthread_should_stop() ||
506 !list_empty(&lo->waiting_queue));
508 /* extract request */
509 if (list_empty(&lo->waiting_queue))
512 spin_lock_irq(&lo->queue_lock);
513 req = list_entry(lo->waiting_queue.next, struct request,
515 list_del_init(&req->queuelist);
516 spin_unlock_irq(&lo->queue_lock);
519 nbd_handle_req(lo, req);
525 * We always wait for result of write, for now. It would be nice to make it optional
527 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
528 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
531 static void do_nbd_request(struct request_queue * q)
535 while ((req = elv_next_request(q)) != NULL) {
536 struct nbd_device *lo;
538 blkdev_dequeue_request(req);
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 spin_lock_irq(&lo->queue_lock);
550 list_add_tail(&req->queuelist, &lo->waiting_queue);
551 spin_unlock_irq(&lo->queue_lock);
553 wake_up(&lo->waiting_wq);
555 spin_lock_irq(q->queue_lock);
559 static int nbd_ioctl(struct inode *inode, struct file *file,
560 unsigned int cmd, unsigned long arg)
562 struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
564 struct request sreq ;
565 struct task_struct *thread;
567 if (!capable(CAP_SYS_ADMIN))
570 BUG_ON(lo->magic != LO_MAGIC);
572 /* Anyone capable of this syscall can do *real bad* things */
573 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
574 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
578 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
579 sreq.cmd_type = REQ_TYPE_SPECIAL;
580 nbd_cmd(&sreq) = NBD_CMD_DISC;
582 * Set these to sane values in case server implementation
583 * fails to check the request type first and also to keep
584 * debugging output cleaner.
590 mutex_lock(&lo->tx_lock);
591 nbd_send_req(lo, &sreq);
592 mutex_unlock(&lo->tx_lock);
597 mutex_lock(&lo->tx_lock);
599 mutex_unlock(&lo->tx_lock);
603 BUG_ON(!list_empty(&lo->queue_head));
613 inode = file->f_path.dentry->d_inode;
614 if (S_ISSOCK(inode->i_mode)) {
616 lo->sock = SOCKET_I(inode);
623 case NBD_SET_BLKSIZE:
625 lo->bytesize &= ~(lo->blksize-1);
626 inode->i_bdev->bd_inode->i_size = lo->bytesize;
627 set_blocksize(inode->i_bdev, lo->blksize);
628 set_capacity(lo->disk, lo->bytesize >> 9);
631 lo->bytesize = arg & ~(lo->blksize-1);
632 inode->i_bdev->bd_inode->i_size = lo->bytesize;
633 set_blocksize(inode->i_bdev, lo->blksize);
634 set_capacity(lo->disk, lo->bytesize >> 9);
636 case NBD_SET_TIMEOUT:
637 lo->xmit_timeout = arg * HZ;
639 case NBD_SET_SIZE_BLOCKS:
640 lo->bytesize = ((u64) arg) * lo->blksize;
641 inode->i_bdev->bd_inode->i_size = lo->bytesize;
642 set_blocksize(inode->i_bdev, lo->blksize);
643 set_capacity(lo->disk, lo->bytesize >> 9);
648 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
650 return PTR_ERR(thread);
651 wake_up_process(thread);
652 error = nbd_do_it(lo);
653 kthread_stop(thread);
656 sock_shutdown(lo, 1);
660 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
664 inode->i_bdev->bd_inode->i_size = 0;
665 set_capacity(lo->disk, 0);
666 return lo->harderror;
669 * This is for compatibility only. The queue is always cleared
670 * by NBD_DO_IT or NBD_CLEAR_SOCK.
672 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
674 case NBD_PRINT_DEBUG:
675 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
676 inode->i_bdev->bd_disk->disk_name,
677 lo->queue_head.next, lo->queue_head.prev,
684 static struct block_device_operations nbd_fops =
686 .owner = THIS_MODULE,
691 * And here should be modules and kernel interface
692 * (Just smiley confuses emacs :-)
695 static int __init nbd_init(void)
700 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
702 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
706 for (i = 0; i < nbds_max; i++) {
707 struct gendisk *disk = alloc_disk(1);
711 nbd_dev[i].disk = disk;
713 * The new linux 2.5 block layer implementation requires
714 * every gendisk to have its very own request_queue struct.
715 * These structs are big so we dynamically allocate them.
717 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
722 old_e = disk->queue->elevator;
723 if (elevator_init(disk->queue, "deadline") == 0 ||
724 elevator_init(disk->queue, "noop") == 0) {
725 elevator_exit(old_e);
729 if (register_blkdev(NBD_MAJOR, "nbd")) {
734 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
735 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
737 for (i = 0; i < nbds_max; i++) {
738 struct gendisk *disk = nbd_dev[i].disk;
739 nbd_dev[i].file = NULL;
740 nbd_dev[i].magic = LO_MAGIC;
741 nbd_dev[i].flags = 0;
742 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
743 spin_lock_init(&nbd_dev[i].queue_lock);
744 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
745 mutex_init(&nbd_dev[i].tx_lock);
746 init_waitqueue_head(&nbd_dev[i].active_wq);
747 init_waitqueue_head(&nbd_dev[i].waiting_wq);
748 nbd_dev[i].blksize = 1024;
749 nbd_dev[i].bytesize = 0;
750 disk->major = NBD_MAJOR;
751 disk->first_minor = i;
752 disk->fops = &nbd_fops;
753 disk->private_data = &nbd_dev[i];
754 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
755 sprintf(disk->disk_name, "nbd%d", i);
756 set_capacity(disk, 0);
763 blk_cleanup_queue(nbd_dev[i].disk->queue);
764 put_disk(nbd_dev[i].disk);
769 static void __exit nbd_cleanup(void)
772 for (i = 0; i < nbds_max; i++) {
773 struct gendisk *disk = nbd_dev[i].disk;
774 nbd_dev[i].magic = 0;
777 blk_cleanup_queue(disk->queue);
781 unregister_blkdev(NBD_MAJOR, "nbd");
782 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
785 module_init(nbd_init);
786 module_exit(nbd_cleanup);
788 MODULE_DESCRIPTION("Network Block Device");
789 MODULE_LICENSE("GPL");
791 module_param(nbds_max, int, 0444);
792 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
794 module_param(debugflags, int, 0644);
795 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");