xen/blkback: fix xenbus_transaction_start() hang caused by double xenbus_transaction_...
[linux-2.6.git] / drivers / block / nbd.c
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
56 static unsigned int debugflags;
57 #endif /* NDEBUG */
58
59 static unsigned int nbds_max = 16;
60 static struct nbd_device *nbd_dev;
61 static 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  */
73 static DEFINE_SPINLOCK(nbd_lock);
74
75 #ifndef NDEBUG
76 static 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
94 static 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
105 static 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
119 static 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
139 static 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  */
151 static 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, 0);
196
197                 if (signal_pending(current)) {
198                         siginfo_t info;
199                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
200                                 task_pid_nr(current), current->comm,
201                                 dequeue_signal_lock(current, &current->blocked, &info));
202                         result = -EINTR;
203                         sock_shutdown(lo, !send);
204                         break;
205                 }
206
207                 if (result <= 0) {
208                         if (result == 0)
209                                 result = -EPIPE; /* short read */
210                         break;
211                 }
212                 size -= result;
213                 buf += result;
214         } while (size > 0);
215
216         sigprocmask(SIG_SETMASK, &oldset, NULL);
217
218         return result;
219 }
220
221 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
222                 int flags)
223 {
224         int result;
225         void *kaddr = kmap(bvec->bv_page);
226         result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
227         kunmap(bvec->bv_page);
228         return result;
229 }
230
231 /* always call with the tx_lock held */
232 static int nbd_send_req(struct nbd_device *lo, struct request *req)
233 {
234         int result, flags;
235         struct nbd_request request;
236         unsigned long size = blk_rq_bytes(req);
237
238         request.magic = htonl(NBD_REQUEST_MAGIC);
239         request.type = htonl(nbd_cmd(req));
240         request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
241         request.len = htonl(size);
242         memcpy(request.handle, &req, sizeof(req));
243
244         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
245                         lo->disk->disk_name, req,
246                         nbdcmd_to_ascii(nbd_cmd(req)),
247                         (unsigned long long)blk_rq_pos(req) << 9,
248                         blk_rq_bytes(req));
249         result = sock_xmit(lo, 1, &request, sizeof(request),
250                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
251         if (result <= 0) {
252                 printk(KERN_ERR "%s: Send control failed (result %d)\n",
253                                 lo->disk->disk_name, result);
254                 goto error_out;
255         }
256
257         if (nbd_cmd(req) == NBD_CMD_WRITE) {
258                 struct req_iterator iter;
259                 struct bio_vec *bvec;
260                 /*
261                  * we are really probing at internals to determine
262                  * whether to set MSG_MORE or not...
263                  */
264                 rq_for_each_segment(bvec, req, iter) {
265                         flags = 0;
266                         if (!rq_iter_last(req, iter))
267                                 flags = MSG_MORE;
268                         dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
269                                         lo->disk->disk_name, req, bvec->bv_len);
270                         result = sock_send_bvec(lo, bvec, flags);
271                         if (result <= 0) {
272                                 printk(KERN_ERR "%s: Send data failed (result %d)\n",
273                                                 lo->disk->disk_name, result);
274                                 goto error_out;
275                         }
276                 }
277         }
278         return 0;
279
280 error_out:
281         return -EIO;
282 }
283
284 static struct request *nbd_find_request(struct nbd_device *lo,
285                                         struct request *xreq)
286 {
287         struct request *req, *tmp;
288         int err;
289
290         err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
291         if (unlikely(err))
292                 goto out;
293
294         spin_lock(&lo->queue_lock);
295         list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
296                 if (req != xreq)
297                         continue;
298                 list_del_init(&req->queuelist);
299                 spin_unlock(&lo->queue_lock);
300                 return req;
301         }
302         spin_unlock(&lo->queue_lock);
303
304         err = -ENOENT;
305
306 out:
307         return ERR_PTR(err);
308 }
309
310 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
311 {
312         int result;
313         void *kaddr = kmap(bvec->bv_page);
314         result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
315                         MSG_WAITALL);
316         kunmap(bvec->bv_page);
317         return result;
318 }
319
320 /* NULL returned = something went wrong, inform userspace */
321 static struct request *nbd_read_stat(struct nbd_device *lo)
322 {
323         int result;
324         struct nbd_reply reply;
325         struct request *req;
326
327         reply.magic = 0;
328         result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
329         if (result <= 0) {
330                 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
331                                 lo->disk->disk_name, result);
332                 goto harderror;
333         }
334
335         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
336                 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
337                                 lo->disk->disk_name,
338                                 (unsigned long)ntohl(reply.magic));
339                 result = -EPROTO;
340                 goto harderror;
341         }
342
343         req = nbd_find_request(lo, *(struct request **)reply.handle);
344         if (IS_ERR(req)) {
345                 result = PTR_ERR(req);
346                 if (result != -ENOENT)
347                         goto harderror;
348
349                 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
350                                 lo->disk->disk_name, reply.handle);
351                 result = -EBADR;
352                 goto harderror;
353         }
354
355         if (ntohl(reply.error)) {
356                 printk(KERN_ERR "%s: Other side returned error (%d)\n",
357                                 lo->disk->disk_name, ntohl(reply.error));
358                 req->errors++;
359                 return req;
360         }
361
362         dprintk(DBG_RX, "%s: request %p: got reply\n",
363                         lo->disk->disk_name, req);
364         if (nbd_cmd(req) == NBD_CMD_READ) {
365                 struct req_iterator iter;
366                 struct bio_vec *bvec;
367
368                 rq_for_each_segment(bvec, req, iter) {
369                         result = sock_recv_bvec(lo, bvec);
370                         if (result <= 0) {
371                                 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
372                                                 lo->disk->disk_name, result);
373                                 req->errors++;
374                                 return req;
375                         }
376                         dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
377                                 lo->disk->disk_name, req, bvec->bv_len);
378                 }
379         }
380         return req;
381 harderror:
382         lo->harderror = result;
383         return NULL;
384 }
385
386 static ssize_t pid_show(struct device *dev,
387                         struct device_attribute *attr, char *buf)
388 {
389         struct gendisk *disk = dev_to_disk(dev);
390
391         return sprintf(buf, "%ld\n",
392                 (long) ((struct nbd_device *)disk->private_data)->pid);
393 }
394
395 static struct device_attribute pid_attr = {
396         .attr = { .name = "pid", .mode = S_IRUGO},
397         .show = pid_show,
398 };
399
400 static int nbd_do_it(struct nbd_device *lo)
401 {
402         struct request *req;
403         int ret;
404
405         BUG_ON(lo->magic != LO_MAGIC);
406
407         lo->pid = current->pid;
408         ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
409         if (ret) {
410                 printk(KERN_ERR "nbd: sysfs_create_file failed!");
411                 lo->pid = 0;
412                 return ret;
413         }
414
415         while ((req = nbd_read_stat(lo)) != NULL)
416                 nbd_end_request(req);
417
418         sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
419         lo->pid = 0;
420         return 0;
421 }
422
423 static void nbd_clear_que(struct nbd_device *lo)
424 {
425         struct request *req;
426
427         BUG_ON(lo->magic != LO_MAGIC);
428
429         /*
430          * Because we have set lo->sock to NULL under the tx_lock, all
431          * modifications to the list must have completed by now.  For
432          * the same reason, the active_req must be NULL.
433          *
434          * As a consequence, we don't need to take the spin lock while
435          * purging the list here.
436          */
437         BUG_ON(lo->sock);
438         BUG_ON(lo->active_req);
439
440         while (!list_empty(&lo->queue_head)) {
441                 req = list_entry(lo->queue_head.next, struct request,
442                                  queuelist);
443                 list_del_init(&req->queuelist);
444                 req->errors++;
445                 nbd_end_request(req);
446         }
447 }
448
449
450 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
451 {
452         if (req->cmd_type != REQ_TYPE_FS)
453                 goto error_out;
454
455         nbd_cmd(req) = NBD_CMD_READ;
456         if (rq_data_dir(req) == WRITE) {
457                 nbd_cmd(req) = NBD_CMD_WRITE;
458                 if (lo->flags & NBD_READ_ONLY) {
459                         printk(KERN_ERR "%s: Write on read-only\n",
460                                         lo->disk->disk_name);
461                         goto error_out;
462                 }
463         }
464
465         req->errors = 0;
466
467         mutex_lock(&lo->tx_lock);
468         if (unlikely(!lo->sock)) {
469                 mutex_unlock(&lo->tx_lock);
470                 printk(KERN_ERR "%s: Attempted send on closed socket\n",
471                        lo->disk->disk_name);
472                 goto error_out;
473         }
474
475         lo->active_req = req;
476
477         if (nbd_send_req(lo, req) != 0) {
478                 printk(KERN_ERR "%s: Request send failed\n",
479                                 lo->disk->disk_name);
480                 req->errors++;
481                 nbd_end_request(req);
482         } else {
483                 spin_lock(&lo->queue_lock);
484                 list_add(&req->queuelist, &lo->queue_head);
485                 spin_unlock(&lo->queue_lock);
486         }
487
488         lo->active_req = NULL;
489         mutex_unlock(&lo->tx_lock);
490         wake_up_all(&lo->active_wq);
491
492         return;
493
494 error_out:
495         req->errors++;
496         nbd_end_request(req);
497 }
498
499 static int nbd_thread(void *data)
500 {
501         struct nbd_device *lo = data;
502         struct request *req;
503
504         set_user_nice(current, -20);
505         while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
506                 /* wait for something to do */
507                 wait_event_interruptible(lo->waiting_wq,
508                                          kthread_should_stop() ||
509                                          !list_empty(&lo->waiting_queue));
510
511                 /* extract request */
512                 if (list_empty(&lo->waiting_queue))
513                         continue;
514
515                 spin_lock_irq(&lo->queue_lock);
516                 req = list_entry(lo->waiting_queue.next, struct request,
517                                  queuelist);
518                 list_del_init(&req->queuelist);
519                 spin_unlock_irq(&lo->queue_lock);
520
521                 /* handle request */
522                 nbd_handle_req(lo, req);
523         }
524         return 0;
525 }
526
527 /*
528  * We always wait for result of write, for now. It would be nice to make it optional
529  * in future
530  * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
531  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
532  */
533
534 static void do_nbd_request(struct request_queue *q)
535 {
536         struct request *req;
537         
538         while ((req = blk_fetch_request(q)) != NULL) {
539                 struct nbd_device *lo;
540
541                 spin_unlock_irq(q->queue_lock);
542
543                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
544                                 req->rq_disk->disk_name, req, req->cmd_type);
545
546                 lo = req->rq_disk->private_data;
547
548                 BUG_ON(lo->magic != LO_MAGIC);
549
550                 if (unlikely(!lo->sock)) {
551                         printk(KERN_ERR "%s: Attempted send on closed socket\n",
552                                 lo->disk->disk_name);
553                         req->errors++;
554                         nbd_end_request(req);
555                         spin_lock_irq(q->queue_lock);
556                         continue;
557                 }
558
559                 spin_lock_irq(&lo->queue_lock);
560                 list_add_tail(&req->queuelist, &lo->waiting_queue);
561                 spin_unlock_irq(&lo->queue_lock);
562
563                 wake_up(&lo->waiting_wq);
564
565                 spin_lock_irq(q->queue_lock);
566         }
567 }
568
569 /* Must be called with tx_lock held */
570
571 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
572                        unsigned int cmd, unsigned long arg)
573 {
574         switch (cmd) {
575         case NBD_DISCONNECT: {
576                 struct request sreq;
577
578                 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
579
580                 blk_rq_init(NULL, &sreq);
581                 sreq.cmd_type = REQ_TYPE_SPECIAL;
582                 nbd_cmd(&sreq) = NBD_CMD_DISC;
583                 if (!lo->sock)
584                         return -EINVAL;
585                 nbd_send_req(lo, &sreq);
586                 return 0;
587         }
588  
589         case NBD_CLEAR_SOCK: {
590                 struct file *file;
591
592                 lo->sock = NULL;
593                 file = lo->file;
594                 lo->file = NULL;
595                 nbd_clear_que(lo);
596                 BUG_ON(!list_empty(&lo->queue_head));
597                 if (file)
598                         fput(file);
599                 return 0;
600         }
601
602         case NBD_SET_SOCK: {
603                 struct file *file;
604                 if (lo->file)
605                         return -EBUSY;
606                 file = fget(arg);
607                 if (file) {
608                         struct inode *inode = file->f_path.dentry->d_inode;
609                         if (S_ISSOCK(inode->i_mode)) {
610                                 lo->file = file;
611                                 lo->sock = SOCKET_I(inode);
612                                 if (max_part > 0)
613                                         bdev->bd_invalidated = 1;
614                                 return 0;
615                         } else {
616                                 fput(file);
617                         }
618                 }
619                 return -EINVAL;
620         }
621
622         case NBD_SET_BLKSIZE:
623                 lo->blksize = arg;
624                 lo->bytesize &= ~(lo->blksize-1);
625                 bdev->bd_inode->i_size = lo->bytesize;
626                 set_blocksize(bdev, lo->blksize);
627                 set_capacity(lo->disk, lo->bytesize >> 9);
628                 return 0;
629
630         case NBD_SET_SIZE:
631                 lo->bytesize = arg & ~(lo->blksize-1);
632                 bdev->bd_inode->i_size = lo->bytesize;
633                 set_blocksize(bdev, lo->blksize);
634                 set_capacity(lo->disk, lo->bytesize >> 9);
635                 return 0;
636
637         case NBD_SET_TIMEOUT:
638                 lo->xmit_timeout = arg * HZ;
639                 return 0;
640
641         case NBD_SET_SIZE_BLOCKS:
642                 lo->bytesize = ((u64) arg) * lo->blksize;
643                 bdev->bd_inode->i_size = lo->bytesize;
644                 set_blocksize(bdev, lo->blksize);
645                 set_capacity(lo->disk, lo->bytesize >> 9);
646                 return 0;
647
648         case NBD_DO_IT: {
649                 struct task_struct *thread;
650                 struct file *file;
651                 int error;
652
653                 if (lo->pid)
654                         return -EBUSY;
655                 if (!lo->file)
656                         return -EINVAL;
657
658                 mutex_unlock(&lo->tx_lock);
659
660                 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
661                 if (IS_ERR(thread)) {
662                         mutex_lock(&lo->tx_lock);
663                         return PTR_ERR(thread);
664                 }
665                 wake_up_process(thread);
666                 error = nbd_do_it(lo);
667                 kthread_stop(thread);
668
669                 mutex_lock(&lo->tx_lock);
670                 if (error)
671                         return error;
672                 sock_shutdown(lo, 0);
673                 file = lo->file;
674                 lo->file = NULL;
675                 nbd_clear_que(lo);
676                 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
677                 if (file)
678                         fput(file);
679                 lo->bytesize = 0;
680                 bdev->bd_inode->i_size = 0;
681                 set_capacity(lo->disk, 0);
682                 if (max_part > 0)
683                         ioctl_by_bdev(bdev, BLKRRPART, 0);
684                 return lo->harderror;
685         }
686
687         case NBD_CLEAR_QUE:
688                 /*
689                  * This is for compatibility only.  The queue is always cleared
690                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
691                  */
692                 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
693                 return 0;
694
695         case NBD_PRINT_DEBUG:
696                 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
697                         bdev->bd_disk->disk_name,
698                         lo->queue_head.next, lo->queue_head.prev,
699                         &lo->queue_head);
700                 return 0;
701         }
702         return -ENOTTY;
703 }
704
705 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
706                      unsigned int cmd, unsigned long arg)
707 {
708         struct nbd_device *lo = bdev->bd_disk->private_data;
709         int error;
710
711         if (!capable(CAP_SYS_ADMIN))
712                 return -EPERM;
713
714         BUG_ON(lo->magic != LO_MAGIC);
715
716         /* Anyone capable of this syscall can do *real bad* things */
717         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
718                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
719
720         mutex_lock(&lo->tx_lock);
721         error = __nbd_ioctl(bdev, lo, cmd, arg);
722         mutex_unlock(&lo->tx_lock);
723
724         return error;
725 }
726
727 static const struct block_device_operations nbd_fops =
728 {
729         .owner =        THIS_MODULE,
730         .ioctl =        nbd_ioctl,
731 };
732
733 /*
734  * And here should be modules and kernel interface 
735  *  (Just smiley confuses emacs :-)
736  */
737
738 static int __init nbd_init(void)
739 {
740         int err = -ENOMEM;
741         int i;
742         int part_shift;
743
744         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
745
746         if (max_part < 0) {
747                 printk(KERN_CRIT "nbd: max_part must be >= 0\n");
748                 return -EINVAL;
749         }
750
751         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
752         if (!nbd_dev)
753                 return -ENOMEM;
754
755         part_shift = 0;
756         if (max_part > 0)
757                 part_shift = fls(max_part);
758
759         for (i = 0; i < nbds_max; i++) {
760                 struct gendisk *disk = alloc_disk(1 << part_shift);
761                 if (!disk)
762                         goto out;
763                 nbd_dev[i].disk = disk;
764                 /*
765                  * The new linux 2.5 block layer implementation requires
766                  * every gendisk to have its very own request_queue struct.
767                  * These structs are big so we dynamically allocate them.
768                  */
769                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
770                 if (!disk->queue) {
771                         put_disk(disk);
772                         goto out;
773                 }
774                 /*
775                  * Tell the block layer that we are not a rotational device
776                  */
777                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
778         }
779
780         if (register_blkdev(NBD_MAJOR, "nbd")) {
781                 err = -EIO;
782                 goto out;
783         }
784
785         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
786         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
787
788         for (i = 0; i < nbds_max; i++) {
789                 struct gendisk *disk = nbd_dev[i].disk;
790                 nbd_dev[i].file = NULL;
791                 nbd_dev[i].magic = LO_MAGIC;
792                 nbd_dev[i].flags = 0;
793                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
794                 spin_lock_init(&nbd_dev[i].queue_lock);
795                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
796                 mutex_init(&nbd_dev[i].tx_lock);
797                 init_waitqueue_head(&nbd_dev[i].active_wq);
798                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
799                 nbd_dev[i].blksize = 1024;
800                 nbd_dev[i].bytesize = 0;
801                 disk->major = NBD_MAJOR;
802                 disk->first_minor = i << part_shift;
803                 disk->fops = &nbd_fops;
804                 disk->private_data = &nbd_dev[i];
805                 sprintf(disk->disk_name, "nbd%d", i);
806                 set_capacity(disk, 0);
807                 add_disk(disk);
808         }
809
810         return 0;
811 out:
812         while (i--) {
813                 blk_cleanup_queue(nbd_dev[i].disk->queue);
814                 put_disk(nbd_dev[i].disk);
815         }
816         kfree(nbd_dev);
817         return err;
818 }
819
820 static void __exit nbd_cleanup(void)
821 {
822         int i;
823         for (i = 0; i < nbds_max; i++) {
824                 struct gendisk *disk = nbd_dev[i].disk;
825                 nbd_dev[i].magic = 0;
826                 if (disk) {
827                         del_gendisk(disk);
828                         blk_cleanup_queue(disk->queue);
829                         put_disk(disk);
830                 }
831         }
832         unregister_blkdev(NBD_MAJOR, "nbd");
833         kfree(nbd_dev);
834         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
835 }
836
837 module_init(nbd_init);
838 module_exit(nbd_cleanup);
839
840 MODULE_DESCRIPTION("Network Block Device");
841 MODULE_LICENSE("GPL");
842
843 module_param(nbds_max, int, 0444);
844 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
845 module_param(max_part, int, 0444);
846 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
847 #ifndef NDEBUG
848 module_param(debugflags, int, 0644);
849 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
850 #endif