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>
32 #include <asm/uaccess.h>
33 #include <asm/system.h>
34 #include <asm/types.h>
36 #include <linux/nbd.h>
38 #define LO_MAGIC 0x68797548
41 #define dprintk(flags, fmt...)
43 #define dprintk(flags, fmt...) do { \
44 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
46 #define DBG_IOCTL 0x0004
47 #define DBG_INIT 0x0010
48 #define DBG_EXIT 0x0020
49 #define DBG_BLKDEV 0x0100
52 static unsigned int debugflags;
55 static unsigned int nbds_max = 16;
56 static struct nbd_device nbd_dev[MAX_NBD];
59 * Use just one lock (or at most 1 per NIC). Two arguments for this:
60 * 1. Each NIC is essentially a synchronization point for all servers
61 * accessed through that NIC so there's no need to have more locks
63 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
64 * down each lock to the point where they're actually slower than just
66 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
68 static DEFINE_SPINLOCK(nbd_lock);
71 static const char *ioctl_cmd_to_ascii(int cmd)
74 case NBD_SET_SOCK: return "set-sock";
75 case NBD_SET_BLKSIZE: return "set-blksize";
76 case NBD_SET_SIZE: return "set-size";
77 case NBD_DO_IT: return "do-it";
78 case NBD_CLEAR_SOCK: return "clear-sock";
79 case NBD_CLEAR_QUE: return "clear-que";
80 case NBD_PRINT_DEBUG: return "print-debug";
81 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
82 case NBD_DISCONNECT: return "disconnect";
83 case BLKROSET: return "set-read-only";
84 case BLKFLSBUF: return "flush-buffer-cache";
89 static const char *nbdcmd_to_ascii(int cmd)
92 case NBD_CMD_READ: return "read";
93 case NBD_CMD_WRITE: return "write";
94 case NBD_CMD_DISC: return "disconnect";
100 static void nbd_end_request(struct request *req)
102 int uptodate = (req->errors == 0) ? 1 : 0;
103 struct request_queue *q = req->q;
106 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
107 req, uptodate? "done": "failed");
109 spin_lock_irqsave(q->queue_lock, flags);
110 if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
111 end_that_request_last(req, uptodate);
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 lo->sock->ops->shutdown(lo->sock, SEND_SHUTDOWN|RCV_SHUTDOWN);
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 /* Allow interception of SIGKILL only
158 * Don't allow other signals to interrupt the transmission */
159 siginitsetinv(&blocked, sigmask(SIGKILL));
160 sigprocmask(SIG_SETMASK, &blocked, &oldset);
163 sock->sk->sk_allocation = GFP_NOIO;
168 msg.msg_control = NULL;
169 msg.msg_controllen = 0;
170 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
173 struct timer_list ti;
175 if (lo->xmit_timeout) {
177 ti.function = nbd_xmit_timeout;
178 ti.data = (unsigned long)current;
179 ti.expires = jiffies + lo->xmit_timeout;
182 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
183 if (lo->xmit_timeout)
186 result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
188 if (signal_pending(current)) {
190 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
191 task_pid_nr(current), current->comm,
192 dequeue_signal_lock(current, ¤t->blocked, &info));
194 sock_shutdown(lo, !send);
200 result = -EPIPE; /* short read */
207 sigprocmask(SIG_SETMASK, &oldset, NULL);
212 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
216 void *kaddr = kmap(bvec->bv_page);
217 result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
218 kunmap(bvec->bv_page);
222 /* always call with the tx_lock held */
223 static int nbd_send_req(struct nbd_device *lo, struct request *req)
226 struct nbd_request request;
227 unsigned long size = req->nr_sectors << 9;
229 request.magic = htonl(NBD_REQUEST_MAGIC);
230 request.type = htonl(nbd_cmd(req));
231 request.from = cpu_to_be64((u64) req->sector << 9);
232 request.len = htonl(size);
233 memcpy(request.handle, &req, sizeof(req));
235 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
236 lo->disk->disk_name, req,
237 nbdcmd_to_ascii(nbd_cmd(req)),
238 (unsigned long long)req->sector << 9,
239 req->nr_sectors << 9);
240 result = sock_xmit(lo, 1, &request, sizeof(request),
241 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
243 printk(KERN_ERR "%s: Send control failed (result %d)\n",
244 lo->disk->disk_name, result);
248 if (nbd_cmd(req) == NBD_CMD_WRITE) {
249 struct req_iterator iter;
250 struct bio_vec *bvec;
252 * we are really probing at internals to determine
253 * whether to set MSG_MORE or not...
255 rq_for_each_segment(bvec, req, iter) {
257 if (!rq_iter_last(req, iter))
259 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
260 lo->disk->disk_name, req, bvec->bv_len);
261 result = sock_send_bvec(lo, bvec, flags);
263 printk(KERN_ERR "%s: Send data failed (result %d)\n",
264 lo->disk->disk_name, result);
275 static struct request *nbd_find_request(struct nbd_device *lo,
276 struct request *xreq)
278 struct request *req, *tmp;
281 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
285 spin_lock(&lo->queue_lock);
286 list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
289 list_del_init(&req->queuelist);
290 spin_unlock(&lo->queue_lock);
293 spin_unlock(&lo->queue_lock);
301 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
304 void *kaddr = kmap(bvec->bv_page);
305 result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
307 kunmap(bvec->bv_page);
311 /* NULL returned = something went wrong, inform userspace */
312 static struct request *nbd_read_stat(struct nbd_device *lo)
315 struct nbd_reply reply;
319 result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
321 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
322 lo->disk->disk_name, result);
326 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
327 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
329 (unsigned long)ntohl(reply.magic));
334 req = nbd_find_request(lo, *(struct request **)reply.handle);
335 if (unlikely(IS_ERR(req))) {
336 result = PTR_ERR(req);
337 if (result != -ENOENT)
340 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
341 lo->disk->disk_name, reply.handle);
346 if (ntohl(reply.error)) {
347 printk(KERN_ERR "%s: Other side returned error (%d)\n",
348 lo->disk->disk_name, ntohl(reply.error));
353 dprintk(DBG_RX, "%s: request %p: got reply\n",
354 lo->disk->disk_name, req);
355 if (nbd_cmd(req) == NBD_CMD_READ) {
356 struct req_iterator iter;
357 struct bio_vec *bvec;
359 rq_for_each_segment(bvec, req, iter) {
360 result = sock_recv_bvec(lo, bvec);
362 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
363 lo->disk->disk_name, result);
367 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
368 lo->disk->disk_name, req, bvec->bv_len);
373 lo->harderror = result;
377 static ssize_t pid_show(struct gendisk *disk, char *page)
379 return sprintf(page, "%ld\n",
380 (long) ((struct nbd_device *)disk->private_data)->pid);
383 static struct disk_attribute pid_attr = {
384 .attr = { .name = "pid", .mode = S_IRUGO },
388 static int nbd_do_it(struct nbd_device *lo)
393 BUG_ON(lo->magic != LO_MAGIC);
395 lo->pid = current->pid;
396 ret = sysfs_create_file(&lo->disk->kobj, &pid_attr.attr);
398 printk(KERN_ERR "nbd: sysfs_create_file failed!");
402 while ((req = nbd_read_stat(lo)) != NULL)
403 nbd_end_request(req);
405 sysfs_remove_file(&lo->disk->kobj, &pid_attr.attr);
409 static void nbd_clear_que(struct nbd_device *lo)
413 BUG_ON(lo->magic != LO_MAGIC);
416 * Because we have set lo->sock to NULL under the tx_lock, all
417 * modifications to the list must have completed by now. For
418 * the same reason, the active_req must be NULL.
420 * As a consequence, we don't need to take the spin lock while
421 * purging the list here.
424 BUG_ON(lo->active_req);
426 while (!list_empty(&lo->queue_head)) {
427 req = list_entry(lo->queue_head.next, struct request,
429 list_del_init(&req->queuelist);
431 nbd_end_request(req);
437 * We always wait for result of write, for now. It would be nice to make it optional
439 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
440 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
443 static void do_nbd_request(struct request_queue * q)
447 while ((req = elv_next_request(q)) != NULL) {
448 struct nbd_device *lo;
450 blkdev_dequeue_request(req);
451 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
452 req->rq_disk->disk_name, req, req->cmd_type);
454 if (!blk_fs_request(req))
457 lo = req->rq_disk->private_data;
459 BUG_ON(lo->magic != LO_MAGIC);
461 nbd_cmd(req) = NBD_CMD_READ;
462 if (rq_data_dir(req) == WRITE) {
463 nbd_cmd(req) = NBD_CMD_WRITE;
464 if (lo->flags & NBD_READ_ONLY) {
465 printk(KERN_ERR "%s: Write on read-only\n",
466 lo->disk->disk_name);
472 spin_unlock_irq(q->queue_lock);
474 mutex_lock(&lo->tx_lock);
475 if (unlikely(!lo->sock)) {
476 mutex_unlock(&lo->tx_lock);
477 printk(KERN_ERR "%s: Attempted send on closed socket\n",
478 lo->disk->disk_name);
480 nbd_end_request(req);
481 spin_lock_irq(q->queue_lock);
485 lo->active_req = req;
487 if (nbd_send_req(lo, req) != 0) {
488 printk(KERN_ERR "%s: Request send failed\n",
489 lo->disk->disk_name);
491 nbd_end_request(req);
493 spin_lock(&lo->queue_lock);
494 list_add(&req->queuelist, &lo->queue_head);
495 spin_unlock(&lo->queue_lock);
498 lo->active_req = NULL;
499 mutex_unlock(&lo->tx_lock);
500 wake_up_all(&lo->active_wq);
502 spin_lock_irq(q->queue_lock);
507 spin_unlock(q->queue_lock);
508 nbd_end_request(req);
509 spin_lock(q->queue_lock);
513 static int nbd_ioctl(struct inode *inode, struct file *file,
514 unsigned int cmd, unsigned long arg)
516 struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
518 struct request sreq ;
520 if (!capable(CAP_SYS_ADMIN))
523 BUG_ON(lo->magic != LO_MAGIC);
525 /* Anyone capable of this syscall can do *real bad* things */
526 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
527 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
531 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
532 sreq.cmd_type = REQ_TYPE_SPECIAL;
533 nbd_cmd(&sreq) = NBD_CMD_DISC;
535 * Set these to sane values in case server implementation
536 * fails to check the request type first and also to keep
537 * debugging output cleaner.
543 mutex_lock(&lo->tx_lock);
544 nbd_send_req(lo, &sreq);
545 mutex_unlock(&lo->tx_lock);
550 mutex_lock(&lo->tx_lock);
552 mutex_unlock(&lo->tx_lock);
556 BUG_ON(!list_empty(&lo->queue_head));
566 inode = file->f_path.dentry->d_inode;
567 if (S_ISSOCK(inode->i_mode)) {
569 lo->sock = SOCKET_I(inode);
576 case NBD_SET_BLKSIZE:
578 lo->bytesize &= ~(lo->blksize-1);
579 inode->i_bdev->bd_inode->i_size = lo->bytesize;
580 set_blocksize(inode->i_bdev, lo->blksize);
581 set_capacity(lo->disk, lo->bytesize >> 9);
584 lo->bytesize = arg & ~(lo->blksize-1);
585 inode->i_bdev->bd_inode->i_size = lo->bytesize;
586 set_blocksize(inode->i_bdev, lo->blksize);
587 set_capacity(lo->disk, lo->bytesize >> 9);
589 case NBD_SET_TIMEOUT:
590 lo->xmit_timeout = arg * HZ;
592 case NBD_SET_SIZE_BLOCKS:
593 lo->bytesize = ((u64) arg) * lo->blksize;
594 inode->i_bdev->bd_inode->i_size = lo->bytesize;
595 set_blocksize(inode->i_bdev, lo->blksize);
596 set_capacity(lo->disk, lo->bytesize >> 9);
601 error = nbd_do_it(lo);
604 sock_shutdown(lo, 1);
608 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
612 inode->i_bdev->bd_inode->i_size = 0;
613 set_capacity(lo->disk, 0);
614 return lo->harderror;
617 * This is for compatibility only. The queue is always cleared
618 * by NBD_DO_IT or NBD_CLEAR_SOCK.
620 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
622 case NBD_PRINT_DEBUG:
623 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
624 inode->i_bdev->bd_disk->disk_name,
625 lo->queue_head.next, lo->queue_head.prev,
632 static struct block_device_operations nbd_fops =
634 .owner = THIS_MODULE,
639 * And here should be modules and kernel interface
640 * (Just smiley confuses emacs :-)
643 static int __init nbd_init(void)
648 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
650 if (nbds_max > MAX_NBD) {
651 printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD,
656 for (i = 0; i < nbds_max; i++) {
657 struct gendisk *disk = alloc_disk(1);
660 nbd_dev[i].disk = disk;
662 * The new linux 2.5 block layer implementation requires
663 * every gendisk to have its very own request_queue struct.
664 * These structs are big so we dynamically allocate them.
666 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
673 if (register_blkdev(NBD_MAJOR, "nbd")) {
678 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
679 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
681 for (i = 0; i < nbds_max; i++) {
682 struct gendisk *disk = nbd_dev[i].disk;
683 nbd_dev[i].file = NULL;
684 nbd_dev[i].magic = LO_MAGIC;
685 nbd_dev[i].flags = 0;
686 spin_lock_init(&nbd_dev[i].queue_lock);
687 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
688 mutex_init(&nbd_dev[i].tx_lock);
689 init_waitqueue_head(&nbd_dev[i].active_wq);
690 nbd_dev[i].blksize = 1024;
691 nbd_dev[i].bytesize = 0;
692 disk->major = NBD_MAJOR;
693 disk->first_minor = i;
694 disk->fops = &nbd_fops;
695 disk->private_data = &nbd_dev[i];
696 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
697 sprintf(disk->disk_name, "nbd%d", i);
698 set_capacity(disk, 0);
705 blk_cleanup_queue(nbd_dev[i].disk->queue);
706 put_disk(nbd_dev[i].disk);
711 static void __exit nbd_cleanup(void)
714 for (i = 0; i < nbds_max; i++) {
715 struct gendisk *disk = nbd_dev[i].disk;
716 nbd_dev[i].magic = 0;
719 blk_cleanup_queue(disk->queue);
723 unregister_blkdev(NBD_MAJOR, "nbd");
724 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
727 module_init(nbd_init);
728 module_exit(nbd_cleanup);
730 MODULE_DESCRIPTION("Network Block Device");
731 MODULE_LICENSE("GPL");
733 module_param(nbds_max, int, 0444);
734 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
736 module_param(debugflags, int, 0644);
737 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");