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;
62 * Use just one lock (or at most 1 per NIC). Two arguments for this:
63 * 1. Each NIC is essentially a synchronization point for all servers
64 * accessed through that NIC so there's no need to have more locks
66 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
67 * down each lock to the point where they're actually slower than just
69 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
71 static DEFINE_SPINLOCK(nbd_lock);
74 static const char *ioctl_cmd_to_ascii(int cmd)
77 case NBD_SET_SOCK: return "set-sock";
78 case NBD_SET_BLKSIZE: return "set-blksize";
79 case NBD_SET_SIZE: return "set-size";
80 case NBD_DO_IT: return "do-it";
81 case NBD_CLEAR_SOCK: return "clear-sock";
82 case NBD_CLEAR_QUE: return "clear-que";
83 case NBD_PRINT_DEBUG: return "print-debug";
84 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
85 case NBD_DISCONNECT: return "disconnect";
86 case BLKROSET: return "set-read-only";
87 case BLKFLSBUF: return "flush-buffer-cache";
92 static const char *nbdcmd_to_ascii(int cmd)
95 case NBD_CMD_READ: return "read";
96 case NBD_CMD_WRITE: return "write";
97 case NBD_CMD_DISC: return "disconnect";
103 static void nbd_end_request(struct request *req)
105 int error = req->errors ? -EIO : 0;
106 struct request_queue *q = req->q;
109 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
110 req, error ? "failed" : "done");
112 spin_lock_irqsave(q->queue_lock, flags);
113 __blk_end_request(req, error, req->nr_sectors << 9);
114 spin_unlock_irqrestore(q->queue_lock, flags);
117 static void sock_shutdown(struct nbd_device *lo, int lock)
119 /* Forcibly shutdown the socket causing all listeners
122 * FIXME: This code is duplicated from sys_shutdown, but
123 * there should be a more generic interface rather than
124 * calling socket ops directly here */
126 mutex_lock(&lo->tx_lock);
128 printk(KERN_WARNING "%s: shutting down socket\n",
129 lo->disk->disk_name);
130 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
134 mutex_unlock(&lo->tx_lock);
137 static void nbd_xmit_timeout(unsigned long arg)
139 struct task_struct *task = (struct task_struct *)arg;
141 printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
142 task->comm, task->pid);
143 force_sig(SIGKILL, task);
147 * Send or receive packet.
149 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
152 struct socket *sock = lo->sock;
156 sigset_t blocked, oldset;
158 if (unlikely(!sock)) {
159 printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
160 lo->disk->disk_name, (send ? "send" : "recv"));
164 /* Allow interception of SIGKILL only
165 * Don't allow other signals to interrupt the transmission */
166 siginitsetinv(&blocked, sigmask(SIGKILL));
167 sigprocmask(SIG_SETMASK, &blocked, &oldset);
170 sock->sk->sk_allocation = GFP_NOIO;
175 msg.msg_control = NULL;
176 msg.msg_controllen = 0;
177 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
180 struct timer_list ti;
182 if (lo->xmit_timeout) {
184 ti.function = nbd_xmit_timeout;
185 ti.data = (unsigned long)current;
186 ti.expires = jiffies + lo->xmit_timeout;
189 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
190 if (lo->xmit_timeout)
193 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
195 if (signal_pending(current)) {
197 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
198 task_pid_nr(current), current->comm,
199 dequeue_signal_lock(current, ¤t->blocked, &info));
201 sock_shutdown(lo, !send);
207 result = -EPIPE; /* short read */
214 sigprocmask(SIG_SETMASK, &oldset, NULL);
219 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
223 void *kaddr = kmap(bvec->bv_page);
224 result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
225 kunmap(bvec->bv_page);
229 /* always call with the tx_lock held */
230 static int nbd_send_req(struct nbd_device *lo, struct request *req)
233 struct nbd_request request;
234 unsigned long size = req->nr_sectors << 9;
236 request.magic = htonl(NBD_REQUEST_MAGIC);
237 request.type = htonl(nbd_cmd(req));
238 request.from = cpu_to_be64((u64) req->sector << 9);
239 request.len = htonl(size);
240 memcpy(request.handle, &req, sizeof(req));
242 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
243 lo->disk->disk_name, req,
244 nbdcmd_to_ascii(nbd_cmd(req)),
245 (unsigned long long)req->sector << 9,
246 req->nr_sectors << 9);
247 result = sock_xmit(lo, 1, &request, sizeof(request),
248 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
250 printk(KERN_ERR "%s: Send control failed (result %d)\n",
251 lo->disk->disk_name, result);
255 if (nbd_cmd(req) == NBD_CMD_WRITE) {
256 struct req_iterator iter;
257 struct bio_vec *bvec;
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(req, iter))
266 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
267 lo->disk->disk_name, req, bvec->bv_len);
268 result = sock_send_bvec(lo, bvec, flags);
270 printk(KERN_ERR "%s: Send data failed (result %d)\n",
271 lo->disk->disk_name, result);
282 static struct request *nbd_find_request(struct nbd_device *lo,
283 struct request *xreq)
285 struct request *req, *tmp;
288 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
292 spin_lock(&lo->queue_lock);
293 list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
296 list_del_init(&req->queuelist);
297 spin_unlock(&lo->queue_lock);
300 spin_unlock(&lo->queue_lock);
308 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
311 void *kaddr = kmap(bvec->bv_page);
312 result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
314 kunmap(bvec->bv_page);
318 /* NULL returned = something went wrong, inform userspace */
319 static struct request *nbd_read_stat(struct nbd_device *lo)
322 struct nbd_reply reply;
326 result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
328 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
329 lo->disk->disk_name, result);
333 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
334 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
336 (unsigned long)ntohl(reply.magic));
341 req = nbd_find_request(lo, *(struct request **)reply.handle);
343 result = PTR_ERR(req);
344 if (result != -ENOENT)
347 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
348 lo->disk->disk_name, reply.handle);
353 if (ntohl(reply.error)) {
354 printk(KERN_ERR "%s: Other side returned error (%d)\n",
355 lo->disk->disk_name, ntohl(reply.error));
360 dprintk(DBG_RX, "%s: request %p: got reply\n",
361 lo->disk->disk_name, req);
362 if (nbd_cmd(req) == NBD_CMD_READ) {
363 struct req_iterator iter;
364 struct bio_vec *bvec;
366 rq_for_each_segment(bvec, req, iter) {
367 result = sock_recv_bvec(lo, bvec);
369 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
370 lo->disk->disk_name, result);
374 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
375 lo->disk->disk_name, req, bvec->bv_len);
380 lo->harderror = result;
384 static ssize_t pid_show(struct device *dev,
385 struct device_attribute *attr, char *buf)
387 struct gendisk *disk = dev_to_disk(dev);
389 return sprintf(buf, "%ld\n",
390 (long) ((struct nbd_device *)disk->private_data)->pid);
393 static struct device_attribute pid_attr = {
394 .attr = { .name = "pid", .mode = S_IRUGO},
398 static int nbd_do_it(struct nbd_device *lo)
403 BUG_ON(lo->magic != LO_MAGIC);
405 lo->pid = current->pid;
406 ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
408 printk(KERN_ERR "nbd: sysfs_create_file failed!");
412 while ((req = nbd_read_stat(lo)) != NULL)
413 nbd_end_request(req);
415 sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
419 static void nbd_clear_que(struct nbd_device *lo)
423 BUG_ON(lo->magic != LO_MAGIC);
426 * Because we have set lo->sock to NULL under the tx_lock, all
427 * modifications to the list must have completed by now. For
428 * the same reason, the active_req must be NULL.
430 * As a consequence, we don't need to take the spin lock while
431 * purging the list here.
434 BUG_ON(lo->active_req);
436 while (!list_empty(&lo->queue_head)) {
437 req = list_entry(lo->queue_head.next, struct request,
439 list_del_init(&req->queuelist);
441 nbd_end_request(req);
446 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
448 if (!blk_fs_request(req))
451 nbd_cmd(req) = NBD_CMD_READ;
452 if (rq_data_dir(req) == WRITE) {
453 nbd_cmd(req) = NBD_CMD_WRITE;
454 if (lo->flags & NBD_READ_ONLY) {
455 printk(KERN_ERR "%s: Write on read-only\n",
456 lo->disk->disk_name);
463 mutex_lock(&lo->tx_lock);
464 if (unlikely(!lo->sock)) {
465 mutex_unlock(&lo->tx_lock);
466 printk(KERN_ERR "%s: Attempted send on closed socket\n",
467 lo->disk->disk_name);
469 nbd_end_request(req);
473 lo->active_req = req;
475 if (nbd_send_req(lo, req) != 0) {
476 printk(KERN_ERR "%s: Request send failed\n",
477 lo->disk->disk_name);
479 nbd_end_request(req);
481 spin_lock(&lo->queue_lock);
482 list_add(&req->queuelist, &lo->queue_head);
483 spin_unlock(&lo->queue_lock);
486 lo->active_req = NULL;
487 mutex_unlock(&lo->tx_lock);
488 wake_up_all(&lo->active_wq);
494 nbd_end_request(req);
497 static int nbd_thread(void *data)
499 struct nbd_device *lo = data;
502 set_user_nice(current, -20);
503 while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
504 /* wait for something to do */
505 wait_event_interruptible(lo->waiting_wq,
506 kthread_should_stop() ||
507 !list_empty(&lo->waiting_queue));
509 /* extract request */
510 if (list_empty(&lo->waiting_queue))
513 spin_lock_irq(&lo->queue_lock);
514 req = list_entry(lo->waiting_queue.next, struct request,
516 list_del_init(&req->queuelist);
517 spin_unlock_irq(&lo->queue_lock);
520 nbd_handle_req(lo, req);
526 * We always wait for result of write, for now. It would be nice to make it optional
528 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
529 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
532 static void do_nbd_request(struct request_queue * q)
536 while ((req = elv_next_request(q)) != NULL) {
537 struct nbd_device *lo;
539 blkdev_dequeue_request(req);
541 spin_unlock_irq(q->queue_lock);
543 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
544 req->rq_disk->disk_name, req, req->cmd_type);
546 lo = req->rq_disk->private_data;
548 BUG_ON(lo->magic != LO_MAGIC);
550 spin_lock_irq(&lo->queue_lock);
551 list_add_tail(&req->queuelist, &lo->waiting_queue);
552 spin_unlock_irq(&lo->queue_lock);
554 wake_up(&lo->waiting_wq);
556 spin_lock_irq(q->queue_lock);
560 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
561 unsigned int cmd, unsigned long arg)
563 struct nbd_device *lo = bdev->bd_disk->private_data;
566 struct request sreq ;
567 struct task_struct *thread;
569 if (!capable(CAP_SYS_ADMIN))
572 BUG_ON(lo->magic != LO_MAGIC);
574 /* Anyone capable of this syscall can do *real bad* things */
575 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
576 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
580 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
581 blk_rq_init(NULL, &sreq);
582 sreq.cmd_type = REQ_TYPE_SPECIAL;
583 nbd_cmd(&sreq) = NBD_CMD_DISC;
585 * Set these to sane values in case server implementation
586 * fails to check the request type first and also to keep
587 * debugging output cleaner.
593 mutex_lock(&lo->tx_lock);
594 nbd_send_req(lo, &sreq);
595 mutex_unlock(&lo->tx_lock);
600 mutex_lock(&lo->tx_lock);
602 mutex_unlock(&lo->tx_lock);
606 BUG_ON(!list_empty(&lo->queue_head));
616 struct inode *inode = file->f_path.dentry->d_inode;
617 if (S_ISSOCK(inode->i_mode)) {
619 lo->sock = SOCKET_I(inode);
621 bdev->bd_invalidated = 1;
628 case NBD_SET_BLKSIZE:
630 lo->bytesize &= ~(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 lo->bytesize = arg & ~(lo->blksize-1);
637 bdev->bd_inode->i_size = lo->bytesize;
638 set_blocksize(bdev, lo->blksize);
639 set_capacity(lo->disk, lo->bytesize >> 9);
641 case NBD_SET_TIMEOUT:
642 lo->xmit_timeout = arg * HZ;
644 case NBD_SET_SIZE_BLOCKS:
645 lo->bytesize = ((u64) arg) * lo->blksize;
646 bdev->bd_inode->i_size = lo->bytesize;
647 set_blocksize(bdev, lo->blksize);
648 set_capacity(lo->disk, lo->bytesize >> 9);
653 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
655 return PTR_ERR(thread);
656 wake_up_process(thread);
657 error = nbd_do_it(lo);
658 kthread_stop(thread);
661 sock_shutdown(lo, 1);
665 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
669 bdev->bd_inode->i_size = 0;
670 set_capacity(lo->disk, 0);
672 ioctl_by_bdev(bdev, BLKRRPART, 0);
673 return lo->harderror;
676 * This is for compatibility only. The queue is always cleared
677 * by NBD_DO_IT or NBD_CLEAR_SOCK.
679 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
681 case NBD_PRINT_DEBUG:
682 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
683 bdev->bd_disk->disk_name,
684 lo->queue_head.next, lo->queue_head.prev,
691 static struct block_device_operations nbd_fops =
693 .owner = THIS_MODULE,
694 .locked_ioctl = nbd_ioctl,
698 * And here should be modules and kernel interface
699 * (Just smiley confuses emacs :-)
702 static int __init nbd_init(void)
708 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
711 printk(KERN_CRIT "nbd: max_part must be >= 0\n");
715 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
721 part_shift = fls(max_part);
723 for (i = 0; i < nbds_max; i++) {
724 struct gendisk *disk = alloc_disk(1 << part_shift);
728 nbd_dev[i].disk = disk;
730 * The new linux 2.5 block layer implementation requires
731 * every gendisk to have its very own request_queue struct.
732 * These structs are big so we dynamically allocate them.
734 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
739 old_e = disk->queue->elevator;
740 if (elevator_init(disk->queue, "deadline") == 0 ||
741 elevator_init(disk->queue, "noop") == 0) {
742 elevator_exit(old_e);
746 if (register_blkdev(NBD_MAJOR, "nbd")) {
751 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
752 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
754 for (i = 0; i < nbds_max; i++) {
755 struct gendisk *disk = nbd_dev[i].disk;
756 nbd_dev[i].file = NULL;
757 nbd_dev[i].magic = LO_MAGIC;
758 nbd_dev[i].flags = 0;
759 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
760 spin_lock_init(&nbd_dev[i].queue_lock);
761 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
762 mutex_init(&nbd_dev[i].tx_lock);
763 init_waitqueue_head(&nbd_dev[i].active_wq);
764 init_waitqueue_head(&nbd_dev[i].waiting_wq);
765 nbd_dev[i].blksize = 1024;
766 nbd_dev[i].bytesize = 0;
767 disk->major = NBD_MAJOR;
768 disk->first_minor = i << part_shift;
769 disk->fops = &nbd_fops;
770 disk->private_data = &nbd_dev[i];
771 sprintf(disk->disk_name, "nbd%d", i);
772 set_capacity(disk, 0);
779 blk_cleanup_queue(nbd_dev[i].disk->queue);
780 put_disk(nbd_dev[i].disk);
786 static void __exit nbd_cleanup(void)
789 for (i = 0; i < nbds_max; i++) {
790 struct gendisk *disk = nbd_dev[i].disk;
791 nbd_dev[i].magic = 0;
794 blk_cleanup_queue(disk->queue);
798 unregister_blkdev(NBD_MAJOR, "nbd");
800 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
803 module_init(nbd_init);
804 module_exit(nbd_cleanup);
806 MODULE_DESCRIPTION("Network Block Device");
807 MODULE_LICENSE("GPL");
809 module_param(nbds_max, int, 0444);
810 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
811 module_param(max_part, int, 0444);
812 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
814 module_param(debugflags, int, 0644);
815 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");