Loading...
1/*
2 * Network block device - make block devices work over TCP
3 *
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.
6 *
7 * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
8 * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
9 *
10 * This file is released under GPLv2 or later.
11 *
12 * (part of code stolen from loop.c)
13 */
14
15#include <linux/major.h>
16
17#include <linux/blkdev.h>
18#include <linux/module.h>
19#include <linux/init.h>
20#include <linux/sched.h>
21#include <linux/fs.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>
32#include <net/sock.h>
33#include <linux/net.h>
34#include <linux/kthread.h>
35
36#include <asm/uaccess.h>
37#include <asm/system.h>
38#include <asm/types.h>
39
40#include <linux/nbd.h>
41
42#define LO_MAGIC 0x68797548
43
44#ifdef NDEBUG
45#define dprintk(flags, fmt...)
46#else /* NDEBUG */
47#define dprintk(flags, fmt...) do { \
48 if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
49} while (0)
50#define DBG_IOCTL 0x0004
51#define DBG_INIT 0x0010
52#define DBG_EXIT 0x0020
53#define DBG_BLKDEV 0x0100
54#define DBG_RX 0x0200
55#define DBG_TX 0x0400
56static unsigned int debugflags;
57#endif /* NDEBUG */
58
59static unsigned int nbds_max = 16;
60static struct nbd_device *nbd_dev;
61static int max_part;
62
63/*
64 * Use just one lock (or at most 1 per NIC). Two arguments for this:
65 * 1. Each NIC is essentially a synchronization point for all servers
66 * accessed through that NIC so there's no need to have more locks
67 * than NICs anyway.
68 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
69 * down each lock to the point where they're actually slower than just
70 * a single lock.
71 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
72 */
73static DEFINE_SPINLOCK(nbd_lock);
74
75#ifndef NDEBUG
76static const char *ioctl_cmd_to_ascii(int cmd)
77{
78 switch (cmd) {
79 case NBD_SET_SOCK: return "set-sock";
80 case NBD_SET_BLKSIZE: return "set-blksize";
81 case NBD_SET_SIZE: return "set-size";
82 case NBD_DO_IT: return "do-it";
83 case NBD_CLEAR_SOCK: return "clear-sock";
84 case NBD_CLEAR_QUE: return "clear-que";
85 case NBD_PRINT_DEBUG: return "print-debug";
86 case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
87 case NBD_DISCONNECT: return "disconnect";
88 case BLKROSET: return "set-read-only";
89 case BLKFLSBUF: return "flush-buffer-cache";
90 }
91 return "unknown";
92}
93
94static const char *nbdcmd_to_ascii(int cmd)
95{
96 switch (cmd) {
97 case NBD_CMD_READ: return "read";
98 case NBD_CMD_WRITE: return "write";
99 case NBD_CMD_DISC: return "disconnect";
100 }
101 return "invalid";
102}
103#endif /* NDEBUG */
104
105static void nbd_end_request(struct request *req)
106{
107 int error = req->errors ? -EIO : 0;
108 struct request_queue *q = req->q;
109 unsigned long flags;
110
111 dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
112 req, error ? "failed" : "done");
113
114 spin_lock_irqsave(q->queue_lock, flags);
115 __blk_end_request_all(req, error);
116 spin_unlock_irqrestore(q->queue_lock, flags);
117}
118
119static void sock_shutdown(struct nbd_device *lo, int lock)
120{
121 /* Forcibly shutdown the socket causing all listeners
122 * to error
123 *
124 * FIXME: This code is duplicated from sys_shutdown, but
125 * there should be a more generic interface rather than
126 * calling socket ops directly here */
127 if (lock)
128 mutex_lock(&lo->tx_lock);
129 if (lo->sock) {
130 printk(KERN_WARNING "%s: shutting down socket\n",
131 lo->disk->disk_name);
132 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
133 lo->sock = NULL;
134 }
135 if (lock)
136 mutex_unlock(&lo->tx_lock);
137}
138
139static void nbd_xmit_timeout(unsigned long arg)
140{
141 struct task_struct *task = (struct task_struct *)arg;
142
143 printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
144 task->comm, task->pid);
145 force_sig(SIGKILL, task);
146}
147
148/*
149 * Send or receive packet.
150 */
151static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
152 int msg_flags)
153{
154 struct socket *sock = lo->sock;
155 int result;
156 struct msghdr msg;
157 struct kvec iov;
158 sigset_t blocked, oldset;
159
160 if (unlikely(!sock)) {
161 printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
162 lo->disk->disk_name, (send ? "send" : "recv"));
163 return -EINVAL;
164 }
165
166 /* Allow interception of SIGKILL only
167 * Don't allow other signals to interrupt the transmission */
168 siginitsetinv(&blocked, sigmask(SIGKILL));
169 sigprocmask(SIG_SETMASK, &blocked, &oldset);
170
171 do {
172 sock->sk->sk_allocation = GFP_NOIO;
173 iov.iov_base = buf;
174 iov.iov_len = size;
175 msg.msg_name = NULL;
176 msg.msg_namelen = 0;
177 msg.msg_control = NULL;
178 msg.msg_controllen = 0;
179 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
180
181 if (send) {
182 struct timer_list ti;
183
184 if (lo->xmit_timeout) {
185 init_timer(&ti);
186 ti.function = nbd_xmit_timeout;
187 ti.data = (unsigned long)current;
188 ti.expires = jiffies + lo->xmit_timeout;
189 add_timer(&ti);
190 }
191 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
192 if (lo->xmit_timeout)
193 del_timer_sync(&ti);
194 } else
195 result = kernel_recvmsg(sock, &msg, &iov, 1, size,
196 msg.msg_flags);
197
198 if (signal_pending(current)) {
199 siginfo_t info;
200 printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
201 task_pid_nr(current), current->comm,
202 dequeue_signal_lock(current, ¤t->blocked, &info));
203 result = -EINTR;
204 sock_shutdown(lo, !send);
205 break;
206 }
207
208 if (result <= 0) {
209 if (result == 0)
210 result = -EPIPE; /* short read */
211 break;
212 }
213 size -= result;
214 buf += result;
215 } while (size > 0);
216
217 sigprocmask(SIG_SETMASK, &oldset, NULL);
218
219 return result;
220}
221
222static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
223 int flags)
224{
225 int result;
226 void *kaddr = kmap(bvec->bv_page);
227 result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
228 kunmap(bvec->bv_page);
229 return result;
230}
231
232/* always call with the tx_lock held */
233static int nbd_send_req(struct nbd_device *lo, struct request *req)
234{
235 int result, flags;
236 struct nbd_request request;
237 unsigned long size = blk_rq_bytes(req);
238
239 request.magic = htonl(NBD_REQUEST_MAGIC);
240 request.type = htonl(nbd_cmd(req));
241 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
242 request.len = htonl(size);
243 memcpy(request.handle, &req, sizeof(req));
244
245 dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
246 lo->disk->disk_name, req,
247 nbdcmd_to_ascii(nbd_cmd(req)),
248 (unsigned long long)blk_rq_pos(req) << 9,
249 blk_rq_bytes(req));
250 result = sock_xmit(lo, 1, &request, sizeof(request),
251 (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
252 if (result <= 0) {
253 printk(KERN_ERR "%s: Send control failed (result %d)\n",
254 lo->disk->disk_name, result);
255 goto error_out;
256 }
257
258 if (nbd_cmd(req) == NBD_CMD_WRITE) {
259 struct req_iterator iter;
260 struct bio_vec *bvec;
261 /*
262 * we are really probing at internals to determine
263 * whether to set MSG_MORE or not...
264 */
265 rq_for_each_segment(bvec, req, iter) {
266 flags = 0;
267 if (!rq_iter_last(req, iter))
268 flags = MSG_MORE;
269 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
270 lo->disk->disk_name, req, bvec->bv_len);
271 result = sock_send_bvec(lo, bvec, flags);
272 if (result <= 0) {
273 printk(KERN_ERR "%s: Send data failed (result %d)\n",
274 lo->disk->disk_name, result);
275 goto error_out;
276 }
277 }
278 }
279 return 0;
280
281error_out:
282 return -EIO;
283}
284
285static struct request *nbd_find_request(struct nbd_device *lo,
286 struct request *xreq)
287{
288 struct request *req, *tmp;
289 int err;
290
291 err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
292 if (unlikely(err))
293 goto out;
294
295 spin_lock(&lo->queue_lock);
296 list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
297 if (req != xreq)
298 continue;
299 list_del_init(&req->queuelist);
300 spin_unlock(&lo->queue_lock);
301 return req;
302 }
303 spin_unlock(&lo->queue_lock);
304
305 err = -ENOENT;
306
307out:
308 return ERR_PTR(err);
309}
310
311static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
312{
313 int result;
314 void *kaddr = kmap(bvec->bv_page);
315 result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
316 MSG_WAITALL);
317 kunmap(bvec->bv_page);
318 return result;
319}
320
321/* NULL returned = something went wrong, inform userspace */
322static struct request *nbd_read_stat(struct nbd_device *lo)
323{
324 int result;
325 struct nbd_reply reply;
326 struct request *req;
327
328 reply.magic = 0;
329 result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
330 if (result <= 0) {
331 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
332 lo->disk->disk_name, result);
333 goto harderror;
334 }
335
336 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
337 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
338 lo->disk->disk_name,
339 (unsigned long)ntohl(reply.magic));
340 result = -EPROTO;
341 goto harderror;
342 }
343
344 req = nbd_find_request(lo, *(struct request **)reply.handle);
345 if (IS_ERR(req)) {
346 result = PTR_ERR(req);
347 if (result != -ENOENT)
348 goto harderror;
349
350 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
351 lo->disk->disk_name, reply.handle);
352 result = -EBADR;
353 goto harderror;
354 }
355
356 if (ntohl(reply.error)) {
357 printk(KERN_ERR "%s: Other side returned error (%d)\n",
358 lo->disk->disk_name, ntohl(reply.error));
359 req->errors++;
360 return req;
361 }
362
363 dprintk(DBG_RX, "%s: request %p: got reply\n",
364 lo->disk->disk_name, req);
365 if (nbd_cmd(req) == NBD_CMD_READ) {
366 struct req_iterator iter;
367 struct bio_vec *bvec;
368
369 rq_for_each_segment(bvec, req, iter) {
370 result = sock_recv_bvec(lo, bvec);
371 if (result <= 0) {
372 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
373 lo->disk->disk_name, result);
374 req->errors++;
375 return req;
376 }
377 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
378 lo->disk->disk_name, req, bvec->bv_len);
379 }
380 }
381 return req;
382harderror:
383 lo->harderror = result;
384 return NULL;
385}
386
387static ssize_t pid_show(struct device *dev,
388 struct device_attribute *attr, char *buf)
389{
390 struct gendisk *disk = dev_to_disk(dev);
391
392 return sprintf(buf, "%ld\n",
393 (long) ((struct nbd_device *)disk->private_data)->pid);
394}
395
396static struct device_attribute pid_attr = {
397 .attr = { .name = "pid", .mode = S_IRUGO},
398 .show = pid_show,
399};
400
401static int nbd_do_it(struct nbd_device *lo)
402{
403 struct request *req;
404 int ret;
405
406 BUG_ON(lo->magic != LO_MAGIC);
407
408 lo->pid = current->pid;
409 ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
410 if (ret) {
411 printk(KERN_ERR "nbd: sysfs_create_file failed!");
412 lo->pid = 0;
413 return ret;
414 }
415
416 while ((req = nbd_read_stat(lo)) != NULL)
417 nbd_end_request(req);
418
419 sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
420 lo->pid = 0;
421 return 0;
422}
423
424static void nbd_clear_que(struct nbd_device *lo)
425{
426 struct request *req;
427
428 BUG_ON(lo->magic != LO_MAGIC);
429
430 /*
431 * Because we have set lo->sock to NULL under the tx_lock, all
432 * modifications to the list must have completed by now. For
433 * the same reason, the active_req must be NULL.
434 *
435 * As a consequence, we don't need to take the spin lock while
436 * purging the list here.
437 */
438 BUG_ON(lo->sock);
439 BUG_ON(lo->active_req);
440
441 while (!list_empty(&lo->queue_head)) {
442 req = list_entry(lo->queue_head.next, struct request,
443 queuelist);
444 list_del_init(&req->queuelist);
445 req->errors++;
446 nbd_end_request(req);
447 }
448}
449
450
451static void nbd_handle_req(struct nbd_device *lo, struct request *req)
452{
453 if (req->cmd_type != REQ_TYPE_FS)
454 goto error_out;
455
456 nbd_cmd(req) = NBD_CMD_READ;
457 if (rq_data_dir(req) == WRITE) {
458 nbd_cmd(req) = NBD_CMD_WRITE;
459 if (lo->flags & NBD_READ_ONLY) {
460 printk(KERN_ERR "%s: Write on read-only\n",
461 lo->disk->disk_name);
462 goto error_out;
463 }
464 }
465
466 req->errors = 0;
467
468 mutex_lock(&lo->tx_lock);
469 if (unlikely(!lo->sock)) {
470 mutex_unlock(&lo->tx_lock);
471 printk(KERN_ERR "%s: Attempted send on closed socket\n",
472 lo->disk->disk_name);
473 goto error_out;
474 }
475
476 lo->active_req = req;
477
478 if (nbd_send_req(lo, req) != 0) {
479 printk(KERN_ERR "%s: Request send failed\n",
480 lo->disk->disk_name);
481 req->errors++;
482 nbd_end_request(req);
483 } else {
484 spin_lock(&lo->queue_lock);
485 list_add(&req->queuelist, &lo->queue_head);
486 spin_unlock(&lo->queue_lock);
487 }
488
489 lo->active_req = NULL;
490 mutex_unlock(&lo->tx_lock);
491 wake_up_all(&lo->active_wq);
492
493 return;
494
495error_out:
496 req->errors++;
497 nbd_end_request(req);
498}
499
500static int nbd_thread(void *data)
501{
502 struct nbd_device *lo = data;
503 struct request *req;
504
505 set_user_nice(current, -20);
506 while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
507 /* wait for something to do */
508 wait_event_interruptible(lo->waiting_wq,
509 kthread_should_stop() ||
510 !list_empty(&lo->waiting_queue));
511
512 /* extract request */
513 if (list_empty(&lo->waiting_queue))
514 continue;
515
516 spin_lock_irq(&lo->queue_lock);
517 req = list_entry(lo->waiting_queue.next, struct request,
518 queuelist);
519 list_del_init(&req->queuelist);
520 spin_unlock_irq(&lo->queue_lock);
521
522 /* handle request */
523 nbd_handle_req(lo, req);
524 }
525 return 0;
526}
527
528/*
529 * We always wait for result of write, for now. It would be nice to make it optional
530 * in future
531 * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
532 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
533 */
534
535static void do_nbd_request(struct request_queue *q)
536{
537 struct request *req;
538
539 while ((req = blk_fetch_request(q)) != NULL) {
540 struct nbd_device *lo;
541
542 spin_unlock_irq(q->queue_lock);
543
544 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
545 req->rq_disk->disk_name, req, req->cmd_type);
546
547 lo = req->rq_disk->private_data;
548
549 BUG_ON(lo->magic != LO_MAGIC);
550
551 if (unlikely(!lo->sock)) {
552 printk(KERN_ERR "%s: Attempted send on closed socket\n",
553 lo->disk->disk_name);
554 req->errors++;
555 nbd_end_request(req);
556 spin_lock_irq(q->queue_lock);
557 continue;
558 }
559
560 spin_lock_irq(&lo->queue_lock);
561 list_add_tail(&req->queuelist, &lo->waiting_queue);
562 spin_unlock_irq(&lo->queue_lock);
563
564 wake_up(&lo->waiting_wq);
565
566 spin_lock_irq(q->queue_lock);
567 }
568}
569
570/* Must be called with tx_lock held */
571
572static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
573 unsigned int cmd, unsigned long arg)
574{
575 switch (cmd) {
576 case NBD_DISCONNECT: {
577 struct request sreq;
578
579 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
580
581 blk_rq_init(NULL, &sreq);
582 sreq.cmd_type = REQ_TYPE_SPECIAL;
583 nbd_cmd(&sreq) = NBD_CMD_DISC;
584 if (!lo->sock)
585 return -EINVAL;
586 nbd_send_req(lo, &sreq);
587 return 0;
588 }
589
590 case NBD_CLEAR_SOCK: {
591 struct file *file;
592
593 lo->sock = NULL;
594 file = lo->file;
595 lo->file = NULL;
596 nbd_clear_que(lo);
597 BUG_ON(!list_empty(&lo->queue_head));
598 if (file)
599 fput(file);
600 return 0;
601 }
602
603 case NBD_SET_SOCK: {
604 struct file *file;
605 if (lo->file)
606 return -EBUSY;
607 file = fget(arg);
608 if (file) {
609 struct inode *inode = file->f_path.dentry->d_inode;
610 if (S_ISSOCK(inode->i_mode)) {
611 lo->file = file;
612 lo->sock = SOCKET_I(inode);
613 if (max_part > 0)
614 bdev->bd_invalidated = 1;
615 return 0;
616 } else {
617 fput(file);
618 }
619 }
620 return -EINVAL;
621 }
622
623 case NBD_SET_BLKSIZE:
624 lo->blksize = arg;
625 lo->bytesize &= ~(lo->blksize-1);
626 bdev->bd_inode->i_size = lo->bytesize;
627 set_blocksize(bdev, lo->blksize);
628 set_capacity(lo->disk, lo->bytesize >> 9);
629 return 0;
630
631 case NBD_SET_SIZE:
632 lo->bytesize = arg & ~(lo->blksize-1);
633 bdev->bd_inode->i_size = lo->bytesize;
634 set_blocksize(bdev, lo->blksize);
635 set_capacity(lo->disk, lo->bytesize >> 9);
636 return 0;
637
638 case NBD_SET_TIMEOUT:
639 lo->xmit_timeout = arg * HZ;
640 return 0;
641
642 case NBD_SET_SIZE_BLOCKS:
643 lo->bytesize = ((u64) arg) * lo->blksize;
644 bdev->bd_inode->i_size = lo->bytesize;
645 set_blocksize(bdev, lo->blksize);
646 set_capacity(lo->disk, lo->bytesize >> 9);
647 return 0;
648
649 case NBD_DO_IT: {
650 struct task_struct *thread;
651 struct file *file;
652 int error;
653
654 if (lo->pid)
655 return -EBUSY;
656 if (!lo->file)
657 return -EINVAL;
658
659 mutex_unlock(&lo->tx_lock);
660
661 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
662 if (IS_ERR(thread)) {
663 mutex_lock(&lo->tx_lock);
664 return PTR_ERR(thread);
665 }
666 wake_up_process(thread);
667 error = nbd_do_it(lo);
668 kthread_stop(thread);
669
670 mutex_lock(&lo->tx_lock);
671 if (error)
672 return error;
673 sock_shutdown(lo, 0);
674 file = lo->file;
675 lo->file = NULL;
676 nbd_clear_que(lo);
677 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
678 if (file)
679 fput(file);
680 lo->bytesize = 0;
681 bdev->bd_inode->i_size = 0;
682 set_capacity(lo->disk, 0);
683 if (max_part > 0)
684 ioctl_by_bdev(bdev, BLKRRPART, 0);
685 return lo->harderror;
686 }
687
688 case NBD_CLEAR_QUE:
689 /*
690 * This is for compatibility only. The queue is always cleared
691 * by NBD_DO_IT or NBD_CLEAR_SOCK.
692 */
693 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
694 return 0;
695
696 case NBD_PRINT_DEBUG:
697 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
698 bdev->bd_disk->disk_name,
699 lo->queue_head.next, lo->queue_head.prev,
700 &lo->queue_head);
701 return 0;
702 }
703 return -ENOTTY;
704}
705
706static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
707 unsigned int cmd, unsigned long arg)
708{
709 struct nbd_device *lo = bdev->bd_disk->private_data;
710 int error;
711
712 if (!capable(CAP_SYS_ADMIN))
713 return -EPERM;
714
715 BUG_ON(lo->magic != LO_MAGIC);
716
717 /* Anyone capable of this syscall can do *real bad* things */
718 dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
719 lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
720
721 mutex_lock(&lo->tx_lock);
722 error = __nbd_ioctl(bdev, lo, cmd, arg);
723 mutex_unlock(&lo->tx_lock);
724
725 return error;
726}
727
728static const struct block_device_operations nbd_fops =
729{
730 .owner = THIS_MODULE,
731 .ioctl = nbd_ioctl,
732};
733
734/*
735 * And here should be modules and kernel interface
736 * (Just smiley confuses emacs :-)
737 */
738
739static int __init nbd_init(void)
740{
741 int err = -ENOMEM;
742 int i;
743 int part_shift;
744
745 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
746
747 if (max_part < 0) {
748 printk(KERN_CRIT "nbd: max_part must be >= 0\n");
749 return -EINVAL;
750 }
751
752 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
753 if (!nbd_dev)
754 return -ENOMEM;
755
756 part_shift = 0;
757 if (max_part > 0) {
758 part_shift = fls(max_part);
759
760 /*
761 * Adjust max_part according to part_shift as it is exported
762 * to user space so that user can know the max number of
763 * partition kernel should be able to manage.
764 *
765 * Note that -1 is required because partition 0 is reserved
766 * for the whole disk.
767 */
768 max_part = (1UL << part_shift) - 1;
769 }
770
771 if ((1UL << part_shift) > DISK_MAX_PARTS)
772 return -EINVAL;
773
774 if (nbds_max > 1UL << (MINORBITS - part_shift))
775 return -EINVAL;
776
777 for (i = 0; i < nbds_max; i++) {
778 struct gendisk *disk = alloc_disk(1 << part_shift);
779 if (!disk)
780 goto out;
781 nbd_dev[i].disk = disk;
782 /*
783 * The new linux 2.5 block layer implementation requires
784 * every gendisk to have its very own request_queue struct.
785 * These structs are big so we dynamically allocate them.
786 */
787 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
788 if (!disk->queue) {
789 put_disk(disk);
790 goto out;
791 }
792 /*
793 * Tell the block layer that we are not a rotational device
794 */
795 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
796 }
797
798 if (register_blkdev(NBD_MAJOR, "nbd")) {
799 err = -EIO;
800 goto out;
801 }
802
803 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
804 dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
805
806 for (i = 0; i < nbds_max; i++) {
807 struct gendisk *disk = nbd_dev[i].disk;
808 nbd_dev[i].file = NULL;
809 nbd_dev[i].magic = LO_MAGIC;
810 nbd_dev[i].flags = 0;
811 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
812 spin_lock_init(&nbd_dev[i].queue_lock);
813 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
814 mutex_init(&nbd_dev[i].tx_lock);
815 init_waitqueue_head(&nbd_dev[i].active_wq);
816 init_waitqueue_head(&nbd_dev[i].waiting_wq);
817 nbd_dev[i].blksize = 1024;
818 nbd_dev[i].bytesize = 0;
819 disk->major = NBD_MAJOR;
820 disk->first_minor = i << part_shift;
821 disk->fops = &nbd_fops;
822 disk->private_data = &nbd_dev[i];
823 sprintf(disk->disk_name, "nbd%d", i);
824 set_capacity(disk, 0);
825 add_disk(disk);
826 }
827
828 return 0;
829out:
830 while (i--) {
831 blk_cleanup_queue(nbd_dev[i].disk->queue);
832 put_disk(nbd_dev[i].disk);
833 }
834 kfree(nbd_dev);
835 return err;
836}
837
838static void __exit nbd_cleanup(void)
839{
840 int i;
841 for (i = 0; i < nbds_max; i++) {
842 struct gendisk *disk = nbd_dev[i].disk;
843 nbd_dev[i].magic = 0;
844 if (disk) {
845 del_gendisk(disk);
846 blk_cleanup_queue(disk->queue);
847 put_disk(disk);
848 }
849 }
850 unregister_blkdev(NBD_MAJOR, "nbd");
851 kfree(nbd_dev);
852 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
853}
854
855module_init(nbd_init);
856module_exit(nbd_cleanup);
857
858MODULE_DESCRIPTION("Network Block Device");
859MODULE_LICENSE("GPL");
860
861module_param(nbds_max, int, 0444);
862MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
863module_param(max_part, int, 0444);
864MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
865#ifndef NDEBUG
866module_param(debugflags, int, 0644);
867MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
868#endif
1/*
2 * Network block device - make block devices work over TCP
3 *
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.
6 *
7 * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
8 * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
9 *
10 * This file is released under GPLv2 or later.
11 *
12 * (part of code stolen from loop.c)
13 */
14
15#include <linux/major.h>
16
17#include <linux/blkdev.h>
18#include <linux/module.h>
19#include <linux/init.h>
20#include <linux/sched.h>
21#include <linux/fs.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>
32#include <net/sock.h>
33#include <linux/net.h>
34#include <linux/kthread.h>
35#include <linux/types.h>
36#include <linux/debugfs.h>
37
38#include <asm/uaccess.h>
39#include <asm/types.h>
40
41#include <linux/nbd.h>
42
43struct nbd_device {
44 u32 flags;
45 struct socket * sock; /* If == NULL, device is not ready, yet */
46 int magic;
47
48 spinlock_t queue_lock;
49 struct list_head queue_head; /* Requests waiting result */
50 struct request *active_req;
51 wait_queue_head_t active_wq;
52 struct list_head waiting_queue; /* Requests to be sent */
53 wait_queue_head_t waiting_wq;
54
55 struct mutex tx_lock;
56 struct gendisk *disk;
57 int blksize;
58 loff_t bytesize;
59 int xmit_timeout;
60 bool timedout;
61 bool disconnect; /* a disconnect has been requested by user */
62
63 struct timer_list timeout_timer;
64 /* protects initialization and shutdown of the socket */
65 spinlock_t sock_lock;
66 struct task_struct *task_recv;
67 struct task_struct *task_send;
68
69#if IS_ENABLED(CONFIG_DEBUG_FS)
70 struct dentry *dbg_dir;
71#endif
72};
73
74#if IS_ENABLED(CONFIG_DEBUG_FS)
75static struct dentry *nbd_dbg_dir;
76#endif
77
78#define nbd_name(nbd) ((nbd)->disk->disk_name)
79
80#define NBD_MAGIC 0x68797548
81
82static unsigned int nbds_max = 16;
83static struct nbd_device *nbd_dev;
84static int max_part;
85
86/*
87 * Use just one lock (or at most 1 per NIC). Two arguments for this:
88 * 1. Each NIC is essentially a synchronization point for all servers
89 * accessed through that NIC so there's no need to have more locks
90 * than NICs anyway.
91 * 2. More locks lead to more "Dirty cache line bouncing" which will slow
92 * down each lock to the point where they're actually slower than just
93 * a single lock.
94 * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
95 */
96static DEFINE_SPINLOCK(nbd_lock);
97
98static inline struct device *nbd_to_dev(struct nbd_device *nbd)
99{
100 return disk_to_dev(nbd->disk);
101}
102
103static bool nbd_is_connected(struct nbd_device *nbd)
104{
105 return !!nbd->task_recv;
106}
107
108static const char *nbdcmd_to_ascii(int cmd)
109{
110 switch (cmd) {
111 case NBD_CMD_READ: return "read";
112 case NBD_CMD_WRITE: return "write";
113 case NBD_CMD_DISC: return "disconnect";
114 case NBD_CMD_FLUSH: return "flush";
115 case NBD_CMD_TRIM: return "trim/discard";
116 }
117 return "invalid";
118}
119
120static int nbd_size_clear(struct nbd_device *nbd, struct block_device *bdev)
121{
122 bdev->bd_inode->i_size = 0;
123 set_capacity(nbd->disk, 0);
124 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
125
126 return 0;
127}
128
129static void nbd_size_update(struct nbd_device *nbd, struct block_device *bdev)
130{
131 if (!nbd_is_connected(nbd))
132 return;
133
134 bdev->bd_inode->i_size = nbd->bytesize;
135 set_capacity(nbd->disk, nbd->bytesize >> 9);
136 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
137}
138
139static int nbd_size_set(struct nbd_device *nbd, struct block_device *bdev,
140 int blocksize, int nr_blocks)
141{
142 int ret;
143
144 ret = set_blocksize(bdev, blocksize);
145 if (ret)
146 return ret;
147
148 nbd->blksize = blocksize;
149 nbd->bytesize = (loff_t)blocksize * (loff_t)nr_blocks;
150
151 nbd_size_update(nbd, bdev);
152
153 return 0;
154}
155
156static void nbd_end_request(struct nbd_device *nbd, struct request *req)
157{
158 int error = req->errors ? -EIO : 0;
159 struct request_queue *q = req->q;
160 unsigned long flags;
161
162 dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
163 error ? "failed" : "done");
164
165 spin_lock_irqsave(q->queue_lock, flags);
166 __blk_end_request_all(req, error);
167 spin_unlock_irqrestore(q->queue_lock, flags);
168}
169
170/*
171 * Forcibly shutdown the socket causing all listeners to error
172 */
173static void sock_shutdown(struct nbd_device *nbd)
174{
175 spin_lock_irq(&nbd->sock_lock);
176
177 if (!nbd->sock) {
178 spin_unlock_irq(&nbd->sock_lock);
179 return;
180 }
181
182 dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
183 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
184 sockfd_put(nbd->sock);
185 nbd->sock = NULL;
186 spin_unlock_irq(&nbd->sock_lock);
187
188 del_timer(&nbd->timeout_timer);
189}
190
191static void nbd_xmit_timeout(unsigned long arg)
192{
193 struct nbd_device *nbd = (struct nbd_device *)arg;
194 unsigned long flags;
195
196 if (list_empty(&nbd->queue_head))
197 return;
198
199 spin_lock_irqsave(&nbd->sock_lock, flags);
200
201 nbd->timedout = true;
202
203 if (nbd->sock)
204 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
205
206 spin_unlock_irqrestore(&nbd->sock_lock, flags);
207
208 dev_err(nbd_to_dev(nbd), "Connection timed out, shutting down connection\n");
209}
210
211/*
212 * Send or receive packet.
213 */
214static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
215 int msg_flags)
216{
217 struct socket *sock = nbd->sock;
218 int result;
219 struct msghdr msg;
220 struct kvec iov;
221 unsigned long pflags = current->flags;
222
223 if (unlikely(!sock)) {
224 dev_err(disk_to_dev(nbd->disk),
225 "Attempted %s on closed socket in sock_xmit\n",
226 (send ? "send" : "recv"));
227 return -EINVAL;
228 }
229
230 current->flags |= PF_MEMALLOC;
231 do {
232 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
233 iov.iov_base = buf;
234 iov.iov_len = size;
235 msg.msg_name = NULL;
236 msg.msg_namelen = 0;
237 msg.msg_control = NULL;
238 msg.msg_controllen = 0;
239 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
240
241 if (send)
242 result = kernel_sendmsg(sock, &msg, &iov, 1, size);
243 else
244 result = kernel_recvmsg(sock, &msg, &iov, 1, size,
245 msg.msg_flags);
246
247 if (result <= 0) {
248 if (result == 0)
249 result = -EPIPE; /* short read */
250 break;
251 }
252 size -= result;
253 buf += result;
254 } while (size > 0);
255
256 tsk_restore_flags(current, pflags, PF_MEMALLOC);
257
258 if (!send && nbd->xmit_timeout)
259 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
260
261 return result;
262}
263
264static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
265 int flags)
266{
267 int result;
268 void *kaddr = kmap(bvec->bv_page);
269 result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
270 bvec->bv_len, flags);
271 kunmap(bvec->bv_page);
272 return result;
273}
274
275/* always call with the tx_lock held */
276static int nbd_send_req(struct nbd_device *nbd, struct request *req)
277{
278 int result, flags;
279 struct nbd_request request;
280 unsigned long size = blk_rq_bytes(req);
281 u32 type;
282
283 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
284 type = NBD_CMD_DISC;
285 else if (req->cmd_flags & REQ_DISCARD)
286 type = NBD_CMD_TRIM;
287 else if (req->cmd_flags & REQ_FLUSH)
288 type = NBD_CMD_FLUSH;
289 else if (rq_data_dir(req) == WRITE)
290 type = NBD_CMD_WRITE;
291 else
292 type = NBD_CMD_READ;
293
294 memset(&request, 0, sizeof(request));
295 request.magic = htonl(NBD_REQUEST_MAGIC);
296 request.type = htonl(type);
297 if (type != NBD_CMD_FLUSH && type != NBD_CMD_DISC) {
298 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
299 request.len = htonl(size);
300 }
301 memcpy(request.handle, &req, sizeof(req));
302
303 dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
304 req, nbdcmd_to_ascii(type),
305 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
306 result = sock_xmit(nbd, 1, &request, sizeof(request),
307 (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
308 if (result <= 0) {
309 dev_err(disk_to_dev(nbd->disk),
310 "Send control failed (result %d)\n", result);
311 return -EIO;
312 }
313
314 if (type == NBD_CMD_WRITE) {
315 struct req_iterator iter;
316 struct bio_vec bvec;
317 /*
318 * we are really probing at internals to determine
319 * whether to set MSG_MORE or not...
320 */
321 rq_for_each_segment(bvec, req, iter) {
322 flags = 0;
323 if (!rq_iter_last(bvec, iter))
324 flags = MSG_MORE;
325 dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
326 req, bvec.bv_len);
327 result = sock_send_bvec(nbd, &bvec, flags);
328 if (result <= 0) {
329 dev_err(disk_to_dev(nbd->disk),
330 "Send data failed (result %d)\n",
331 result);
332 return -EIO;
333 }
334 }
335 }
336 return 0;
337}
338
339static struct request *nbd_find_request(struct nbd_device *nbd,
340 struct request *xreq)
341{
342 struct request *req, *tmp;
343 int err;
344
345 err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
346 if (unlikely(err))
347 return ERR_PTR(err);
348
349 spin_lock(&nbd->queue_lock);
350 list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
351 if (req != xreq)
352 continue;
353 list_del_init(&req->queuelist);
354 spin_unlock(&nbd->queue_lock);
355 return req;
356 }
357 spin_unlock(&nbd->queue_lock);
358
359 return ERR_PTR(-ENOENT);
360}
361
362static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
363{
364 int result;
365 void *kaddr = kmap(bvec->bv_page);
366 result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
367 MSG_WAITALL);
368 kunmap(bvec->bv_page);
369 return result;
370}
371
372/* NULL returned = something went wrong, inform userspace */
373static struct request *nbd_read_stat(struct nbd_device *nbd)
374{
375 int result;
376 struct nbd_reply reply;
377 struct request *req;
378
379 reply.magic = 0;
380 result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
381 if (result <= 0) {
382 dev_err(disk_to_dev(nbd->disk),
383 "Receive control failed (result %d)\n", result);
384 return ERR_PTR(result);
385 }
386
387 if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
388 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
389 (unsigned long)ntohl(reply.magic));
390 return ERR_PTR(-EPROTO);
391 }
392
393 req = nbd_find_request(nbd, *(struct request **)reply.handle);
394 if (IS_ERR(req)) {
395 result = PTR_ERR(req);
396 if (result != -ENOENT)
397 return ERR_PTR(result);
398
399 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
400 reply.handle);
401 return ERR_PTR(-EBADR);
402 }
403
404 if (ntohl(reply.error)) {
405 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
406 ntohl(reply.error));
407 req->errors++;
408 return req;
409 }
410
411 dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
412 if (rq_data_dir(req) != WRITE) {
413 struct req_iterator iter;
414 struct bio_vec bvec;
415
416 rq_for_each_segment(bvec, req, iter) {
417 result = sock_recv_bvec(nbd, &bvec);
418 if (result <= 0) {
419 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
420 result);
421 req->errors++;
422 return req;
423 }
424 dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
425 req, bvec.bv_len);
426 }
427 }
428 return req;
429}
430
431static ssize_t pid_show(struct device *dev,
432 struct device_attribute *attr, char *buf)
433{
434 struct gendisk *disk = dev_to_disk(dev);
435 struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
436
437 return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
438}
439
440static struct device_attribute pid_attr = {
441 .attr = { .name = "pid", .mode = S_IRUGO},
442 .show = pid_show,
443};
444
445static int nbd_thread_recv(struct nbd_device *nbd, struct block_device *bdev)
446{
447 struct request *req;
448 int ret;
449
450 BUG_ON(nbd->magic != NBD_MAGIC);
451
452 sk_set_memalloc(nbd->sock->sk);
453
454 nbd->task_recv = current;
455
456 ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
457 if (ret) {
458 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
459
460 nbd->task_recv = NULL;
461
462 return ret;
463 }
464
465 nbd_size_update(nbd, bdev);
466
467 while (1) {
468 req = nbd_read_stat(nbd);
469 if (IS_ERR(req)) {
470 ret = PTR_ERR(req);
471 break;
472 }
473
474 nbd_end_request(nbd, req);
475 }
476
477 nbd_size_clear(nbd, bdev);
478
479 device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
480
481 nbd->task_recv = NULL;
482
483 return ret;
484}
485
486static void nbd_clear_que(struct nbd_device *nbd)
487{
488 struct request *req;
489
490 BUG_ON(nbd->magic != NBD_MAGIC);
491
492 /*
493 * Because we have set nbd->sock to NULL under the tx_lock, all
494 * modifications to the list must have completed by now. For
495 * the same reason, the active_req must be NULL.
496 *
497 * As a consequence, we don't need to take the spin lock while
498 * purging the list here.
499 */
500 BUG_ON(nbd->sock);
501 BUG_ON(nbd->active_req);
502
503 while (!list_empty(&nbd->queue_head)) {
504 req = list_entry(nbd->queue_head.next, struct request,
505 queuelist);
506 list_del_init(&req->queuelist);
507 req->errors++;
508 nbd_end_request(nbd, req);
509 }
510
511 while (!list_empty(&nbd->waiting_queue)) {
512 req = list_entry(nbd->waiting_queue.next, struct request,
513 queuelist);
514 list_del_init(&req->queuelist);
515 req->errors++;
516 nbd_end_request(nbd, req);
517 }
518 dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
519}
520
521
522static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
523{
524 if (req->cmd_type != REQ_TYPE_FS)
525 goto error_out;
526
527 if (rq_data_dir(req) == WRITE &&
528 (nbd->flags & NBD_FLAG_READ_ONLY)) {
529 dev_err(disk_to_dev(nbd->disk),
530 "Write on read-only\n");
531 goto error_out;
532 }
533
534 req->errors = 0;
535
536 mutex_lock(&nbd->tx_lock);
537 if (unlikely(!nbd->sock)) {
538 mutex_unlock(&nbd->tx_lock);
539 dev_err(disk_to_dev(nbd->disk),
540 "Attempted send on closed socket\n");
541 goto error_out;
542 }
543
544 nbd->active_req = req;
545
546 if (nbd->xmit_timeout && list_empty_careful(&nbd->queue_head))
547 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
548
549 if (nbd_send_req(nbd, req) != 0) {
550 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
551 req->errors++;
552 nbd_end_request(nbd, req);
553 } else {
554 spin_lock(&nbd->queue_lock);
555 list_add_tail(&req->queuelist, &nbd->queue_head);
556 spin_unlock(&nbd->queue_lock);
557 }
558
559 nbd->active_req = NULL;
560 mutex_unlock(&nbd->tx_lock);
561 wake_up_all(&nbd->active_wq);
562
563 return;
564
565error_out:
566 req->errors++;
567 nbd_end_request(nbd, req);
568}
569
570static int nbd_thread_send(void *data)
571{
572 struct nbd_device *nbd = data;
573 struct request *req;
574
575 nbd->task_send = current;
576
577 set_user_nice(current, MIN_NICE);
578 while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
579 /* wait for something to do */
580 wait_event_interruptible(nbd->waiting_wq,
581 kthread_should_stop() ||
582 !list_empty(&nbd->waiting_queue));
583
584 /* extract request */
585 if (list_empty(&nbd->waiting_queue))
586 continue;
587
588 spin_lock_irq(&nbd->queue_lock);
589 req = list_entry(nbd->waiting_queue.next, struct request,
590 queuelist);
591 list_del_init(&req->queuelist);
592 spin_unlock_irq(&nbd->queue_lock);
593
594 /* handle request */
595 nbd_handle_req(nbd, req);
596 }
597
598 nbd->task_send = NULL;
599
600 return 0;
601}
602
603/*
604 * We always wait for result of write, for now. It would be nice to make it optional
605 * in future
606 * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
607 * { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
608 */
609
610static void nbd_request_handler(struct request_queue *q)
611 __releases(q->queue_lock) __acquires(q->queue_lock)
612{
613 struct request *req;
614
615 while ((req = blk_fetch_request(q)) != NULL) {
616 struct nbd_device *nbd;
617
618 spin_unlock_irq(q->queue_lock);
619
620 nbd = req->rq_disk->private_data;
621
622 BUG_ON(nbd->magic != NBD_MAGIC);
623
624 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
625 req, req->cmd_type);
626
627 if (unlikely(!nbd->sock)) {
628 dev_err_ratelimited(disk_to_dev(nbd->disk),
629 "Attempted send on closed socket\n");
630 req->errors++;
631 nbd_end_request(nbd, req);
632 spin_lock_irq(q->queue_lock);
633 continue;
634 }
635
636 spin_lock_irq(&nbd->queue_lock);
637 list_add_tail(&req->queuelist, &nbd->waiting_queue);
638 spin_unlock_irq(&nbd->queue_lock);
639
640 wake_up(&nbd->waiting_wq);
641
642 spin_lock_irq(q->queue_lock);
643 }
644}
645
646static int nbd_set_socket(struct nbd_device *nbd, struct socket *sock)
647{
648 int ret = 0;
649
650 spin_lock_irq(&nbd->sock_lock);
651
652 if (nbd->sock) {
653 ret = -EBUSY;
654 goto out;
655 }
656
657 nbd->sock = sock;
658
659out:
660 spin_unlock_irq(&nbd->sock_lock);
661
662 return ret;
663}
664
665/* Reset all properties of an NBD device */
666static void nbd_reset(struct nbd_device *nbd)
667{
668 nbd->disconnect = false;
669 nbd->timedout = false;
670 nbd->blksize = 1024;
671 nbd->bytesize = 0;
672 set_capacity(nbd->disk, 0);
673 nbd->flags = 0;
674 nbd->xmit_timeout = 0;
675 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
676 del_timer_sync(&nbd->timeout_timer);
677}
678
679static void nbd_bdev_reset(struct block_device *bdev)
680{
681 set_device_ro(bdev, false);
682 bdev->bd_inode->i_size = 0;
683 if (max_part > 0) {
684 blkdev_reread_part(bdev);
685 bdev->bd_invalidated = 1;
686 }
687}
688
689static void nbd_parse_flags(struct nbd_device *nbd, struct block_device *bdev)
690{
691 if (nbd->flags & NBD_FLAG_READ_ONLY)
692 set_device_ro(bdev, true);
693 if (nbd->flags & NBD_FLAG_SEND_TRIM)
694 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
695 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
696 blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
697 else
698 blk_queue_flush(nbd->disk->queue, 0);
699}
700
701static int nbd_dev_dbg_init(struct nbd_device *nbd);
702static void nbd_dev_dbg_close(struct nbd_device *nbd);
703
704/* Must be called with tx_lock held */
705
706static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
707 unsigned int cmd, unsigned long arg)
708{
709 switch (cmd) {
710 case NBD_DISCONNECT: {
711 struct request sreq;
712
713 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
714 if (!nbd->sock)
715 return -EINVAL;
716
717 mutex_unlock(&nbd->tx_lock);
718 fsync_bdev(bdev);
719 mutex_lock(&nbd->tx_lock);
720 blk_rq_init(NULL, &sreq);
721 sreq.cmd_type = REQ_TYPE_DRV_PRIV;
722
723 /* Check again after getting mutex back. */
724 if (!nbd->sock)
725 return -EINVAL;
726
727 nbd->disconnect = true;
728
729 nbd_send_req(nbd, &sreq);
730 return 0;
731 }
732
733 case NBD_CLEAR_SOCK:
734 sock_shutdown(nbd);
735 nbd_clear_que(nbd);
736 BUG_ON(!list_empty(&nbd->queue_head));
737 BUG_ON(!list_empty(&nbd->waiting_queue));
738 kill_bdev(bdev);
739 return 0;
740
741 case NBD_SET_SOCK: {
742 int err;
743 struct socket *sock = sockfd_lookup(arg, &err);
744
745 if (!sock)
746 return err;
747
748 err = nbd_set_socket(nbd, sock);
749 if (!err && max_part)
750 bdev->bd_invalidated = 1;
751
752 return err;
753 }
754
755 case NBD_SET_BLKSIZE: {
756 loff_t bsize = div_s64(nbd->bytesize, arg);
757
758 return nbd_size_set(nbd, bdev, arg, bsize);
759 }
760
761 case NBD_SET_SIZE:
762 return nbd_size_set(nbd, bdev, nbd->blksize,
763 arg / nbd->blksize);
764
765 case NBD_SET_SIZE_BLOCKS:
766 return nbd_size_set(nbd, bdev, nbd->blksize, arg);
767
768 case NBD_SET_TIMEOUT:
769 nbd->xmit_timeout = arg * HZ;
770 if (arg)
771 mod_timer(&nbd->timeout_timer,
772 jiffies + nbd->xmit_timeout);
773 else
774 del_timer_sync(&nbd->timeout_timer);
775
776 return 0;
777
778 case NBD_SET_FLAGS:
779 nbd->flags = arg;
780 return 0;
781
782 case NBD_DO_IT: {
783 struct task_struct *thread;
784 int error;
785
786 if (nbd->task_recv)
787 return -EBUSY;
788 if (!nbd->sock)
789 return -EINVAL;
790
791 mutex_unlock(&nbd->tx_lock);
792
793 nbd_parse_flags(nbd, bdev);
794
795 thread = kthread_run(nbd_thread_send, nbd, "%s",
796 nbd_name(nbd));
797 if (IS_ERR(thread)) {
798 mutex_lock(&nbd->tx_lock);
799 return PTR_ERR(thread);
800 }
801
802 nbd_dev_dbg_init(nbd);
803 error = nbd_thread_recv(nbd, bdev);
804 nbd_dev_dbg_close(nbd);
805 kthread_stop(thread);
806
807 mutex_lock(&nbd->tx_lock);
808
809 sock_shutdown(nbd);
810 nbd_clear_que(nbd);
811 kill_bdev(bdev);
812 nbd_bdev_reset(bdev);
813
814 if (nbd->disconnect) /* user requested, ignore socket errors */
815 error = 0;
816 if (nbd->timedout)
817 error = -ETIMEDOUT;
818
819 nbd_reset(nbd);
820
821 return error;
822 }
823
824 case NBD_CLEAR_QUE:
825 /*
826 * This is for compatibility only. The queue is always cleared
827 * by NBD_DO_IT or NBD_CLEAR_SOCK.
828 */
829 return 0;
830
831 case NBD_PRINT_DEBUG:
832 dev_info(disk_to_dev(nbd->disk),
833 "next = %p, prev = %p, head = %p\n",
834 nbd->queue_head.next, nbd->queue_head.prev,
835 &nbd->queue_head);
836 return 0;
837 }
838 return -ENOTTY;
839}
840
841static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
842 unsigned int cmd, unsigned long arg)
843{
844 struct nbd_device *nbd = bdev->bd_disk->private_data;
845 int error;
846
847 if (!capable(CAP_SYS_ADMIN))
848 return -EPERM;
849
850 BUG_ON(nbd->magic != NBD_MAGIC);
851
852 mutex_lock(&nbd->tx_lock);
853 error = __nbd_ioctl(bdev, nbd, cmd, arg);
854 mutex_unlock(&nbd->tx_lock);
855
856 return error;
857}
858
859static const struct block_device_operations nbd_fops =
860{
861 .owner = THIS_MODULE,
862 .ioctl = nbd_ioctl,
863 .compat_ioctl = nbd_ioctl,
864};
865
866#if IS_ENABLED(CONFIG_DEBUG_FS)
867
868static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
869{
870 struct nbd_device *nbd = s->private;
871
872 if (nbd->task_recv)
873 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
874 if (nbd->task_send)
875 seq_printf(s, "send: %d\n", task_pid_nr(nbd->task_send));
876
877 return 0;
878}
879
880static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
881{
882 return single_open(file, nbd_dbg_tasks_show, inode->i_private);
883}
884
885static const struct file_operations nbd_dbg_tasks_ops = {
886 .open = nbd_dbg_tasks_open,
887 .read = seq_read,
888 .llseek = seq_lseek,
889 .release = single_release,
890};
891
892static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
893{
894 struct nbd_device *nbd = s->private;
895 u32 flags = nbd->flags;
896
897 seq_printf(s, "Hex: 0x%08x\n\n", flags);
898
899 seq_puts(s, "Known flags:\n");
900
901 if (flags & NBD_FLAG_HAS_FLAGS)
902 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
903 if (flags & NBD_FLAG_READ_ONLY)
904 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
905 if (flags & NBD_FLAG_SEND_FLUSH)
906 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
907 if (flags & NBD_FLAG_SEND_TRIM)
908 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
909
910 return 0;
911}
912
913static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
914{
915 return single_open(file, nbd_dbg_flags_show, inode->i_private);
916}
917
918static const struct file_operations nbd_dbg_flags_ops = {
919 .open = nbd_dbg_flags_open,
920 .read = seq_read,
921 .llseek = seq_lseek,
922 .release = single_release,
923};
924
925static int nbd_dev_dbg_init(struct nbd_device *nbd)
926{
927 struct dentry *dir;
928
929 if (!nbd_dbg_dir)
930 return -EIO;
931
932 dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
933 if (!dir) {
934 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
935 nbd_name(nbd));
936 return -EIO;
937 }
938 nbd->dbg_dir = dir;
939
940 debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
941 debugfs_create_u64("size_bytes", 0444, dir, &nbd->bytesize);
942 debugfs_create_u32("timeout", 0444, dir, &nbd->xmit_timeout);
943 debugfs_create_u32("blocksize", 0444, dir, &nbd->blksize);
944 debugfs_create_file("flags", 0444, dir, &nbd, &nbd_dbg_flags_ops);
945
946 return 0;
947}
948
949static void nbd_dev_dbg_close(struct nbd_device *nbd)
950{
951 debugfs_remove_recursive(nbd->dbg_dir);
952}
953
954static int nbd_dbg_init(void)
955{
956 struct dentry *dbg_dir;
957
958 dbg_dir = debugfs_create_dir("nbd", NULL);
959 if (!dbg_dir)
960 return -EIO;
961
962 nbd_dbg_dir = dbg_dir;
963
964 return 0;
965}
966
967static void nbd_dbg_close(void)
968{
969 debugfs_remove_recursive(nbd_dbg_dir);
970}
971
972#else /* IS_ENABLED(CONFIG_DEBUG_FS) */
973
974static int nbd_dev_dbg_init(struct nbd_device *nbd)
975{
976 return 0;
977}
978
979static void nbd_dev_dbg_close(struct nbd_device *nbd)
980{
981}
982
983static int nbd_dbg_init(void)
984{
985 return 0;
986}
987
988static void nbd_dbg_close(void)
989{
990}
991
992#endif
993
994/*
995 * And here should be modules and kernel interface
996 * (Just smiley confuses emacs :-)
997 */
998
999static int __init nbd_init(void)
1000{
1001 int err = -ENOMEM;
1002 int i;
1003 int part_shift;
1004
1005 BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
1006
1007 if (max_part < 0) {
1008 printk(KERN_ERR "nbd: max_part must be >= 0\n");
1009 return -EINVAL;
1010 }
1011
1012 part_shift = 0;
1013 if (max_part > 0) {
1014 part_shift = fls(max_part);
1015
1016 /*
1017 * Adjust max_part according to part_shift as it is exported
1018 * to user space so that user can know the max number of
1019 * partition kernel should be able to manage.
1020 *
1021 * Note that -1 is required because partition 0 is reserved
1022 * for the whole disk.
1023 */
1024 max_part = (1UL << part_shift) - 1;
1025 }
1026
1027 if ((1UL << part_shift) > DISK_MAX_PARTS)
1028 return -EINVAL;
1029
1030 if (nbds_max > 1UL << (MINORBITS - part_shift))
1031 return -EINVAL;
1032
1033 nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
1034 if (!nbd_dev)
1035 return -ENOMEM;
1036
1037 for (i = 0; i < nbds_max; i++) {
1038 struct gendisk *disk = alloc_disk(1 << part_shift);
1039 if (!disk)
1040 goto out;
1041 nbd_dev[i].disk = disk;
1042 /*
1043 * The new linux 2.5 block layer implementation requires
1044 * every gendisk to have its very own request_queue struct.
1045 * These structs are big so we dynamically allocate them.
1046 */
1047 disk->queue = blk_init_queue(nbd_request_handler, &nbd_lock);
1048 if (!disk->queue) {
1049 put_disk(disk);
1050 goto out;
1051 }
1052 /*
1053 * Tell the block layer that we are not a rotational device
1054 */
1055 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
1056 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1057 disk->queue->limits.discard_granularity = 512;
1058 blk_queue_max_discard_sectors(disk->queue, UINT_MAX);
1059 disk->queue->limits.discard_zeroes_data = 0;
1060 blk_queue_max_hw_sectors(disk->queue, 65536);
1061 disk->queue->limits.max_sectors = 256;
1062 }
1063
1064 if (register_blkdev(NBD_MAJOR, "nbd")) {
1065 err = -EIO;
1066 goto out;
1067 }
1068
1069 printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
1070
1071 nbd_dbg_init();
1072
1073 for (i = 0; i < nbds_max; i++) {
1074 struct gendisk *disk = nbd_dev[i].disk;
1075 nbd_dev[i].magic = NBD_MAGIC;
1076 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
1077 spin_lock_init(&nbd_dev[i].queue_lock);
1078 spin_lock_init(&nbd_dev[i].sock_lock);
1079 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
1080 mutex_init(&nbd_dev[i].tx_lock);
1081 init_timer(&nbd_dev[i].timeout_timer);
1082 nbd_dev[i].timeout_timer.function = nbd_xmit_timeout;
1083 nbd_dev[i].timeout_timer.data = (unsigned long)&nbd_dev[i];
1084 init_waitqueue_head(&nbd_dev[i].active_wq);
1085 init_waitqueue_head(&nbd_dev[i].waiting_wq);
1086 disk->major = NBD_MAJOR;
1087 disk->first_minor = i << part_shift;
1088 disk->fops = &nbd_fops;
1089 disk->private_data = &nbd_dev[i];
1090 sprintf(disk->disk_name, "nbd%d", i);
1091 nbd_reset(&nbd_dev[i]);
1092 add_disk(disk);
1093 }
1094
1095 return 0;
1096out:
1097 while (i--) {
1098 blk_cleanup_queue(nbd_dev[i].disk->queue);
1099 put_disk(nbd_dev[i].disk);
1100 }
1101 kfree(nbd_dev);
1102 return err;
1103}
1104
1105static void __exit nbd_cleanup(void)
1106{
1107 int i;
1108
1109 nbd_dbg_close();
1110
1111 for (i = 0; i < nbds_max; i++) {
1112 struct gendisk *disk = nbd_dev[i].disk;
1113 nbd_dev[i].magic = 0;
1114 if (disk) {
1115 del_gendisk(disk);
1116 blk_cleanup_queue(disk->queue);
1117 put_disk(disk);
1118 }
1119 }
1120 unregister_blkdev(NBD_MAJOR, "nbd");
1121 kfree(nbd_dev);
1122 printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
1123}
1124
1125module_init(nbd_init);
1126module_exit(nbd_cleanup);
1127
1128MODULE_DESCRIPTION("Network Block Device");
1129MODULE_LICENSE("GPL");
1130
1131module_param(nbds_max, int, 0444);
1132MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
1133module_param(max_part, int, 0444);
1134MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");