xref: /linux/block/bsg.c (revision ecba1060583635ab55092072441ff903b5e9a659)
1 /*
2  * bsg.c - block layer implementation of the sg v4 interface
3  *
4  * Copyright (C) 2004 Jens Axboe <axboe@suse.de> SUSE Labs
5  * Copyright (C) 2004 Peter M. Jones <pjones@redhat.com>
6  *
7  *  This file is subject to the terms and conditions of the GNU General Public
8  *  License version 2.  See the file "COPYING" in the main directory of this
9  *  archive for more details.
10  *
11  */
12 #include <linux/module.h>
13 #include <linux/init.h>
14 #include <linux/file.h>
15 #include <linux/blkdev.h>
16 #include <linux/poll.h>
17 #include <linux/cdev.h>
18 #include <linux/percpu.h>
19 #include <linux/uio.h>
20 #include <linux/idr.h>
21 #include <linux/bsg.h>
22 #include <linux/smp_lock.h>
23 
24 #include <scsi/scsi.h>
25 #include <scsi/scsi_ioctl.h>
26 #include <scsi/scsi_cmnd.h>
27 #include <scsi/scsi_device.h>
28 #include <scsi/scsi_driver.h>
29 #include <scsi/sg.h>
30 
31 #define BSG_DESCRIPTION	"Block layer SCSI generic (bsg) driver"
32 #define BSG_VERSION	"0.4"
33 
34 struct bsg_device {
35 	struct request_queue *queue;
36 	spinlock_t lock;
37 	struct list_head busy_list;
38 	struct list_head done_list;
39 	struct hlist_node dev_list;
40 	atomic_t ref_count;
41 	int queued_cmds;
42 	int done_cmds;
43 	wait_queue_head_t wq_done;
44 	wait_queue_head_t wq_free;
45 	char name[20];
46 	int max_queue;
47 	unsigned long flags;
48 };
49 
50 enum {
51 	BSG_F_BLOCK		= 1,
52 };
53 
54 #define BSG_DEFAULT_CMDS	64
55 #define BSG_MAX_DEVS		32768
56 
57 #undef BSG_DEBUG
58 
59 #ifdef BSG_DEBUG
60 #define dprintk(fmt, args...) printk(KERN_ERR "%s: " fmt, __func__, ##args)
61 #else
62 #define dprintk(fmt, args...)
63 #endif
64 
65 static DEFINE_MUTEX(bsg_mutex);
66 static DEFINE_IDR(bsg_minor_idr);
67 
68 #define BSG_LIST_ARRAY_SIZE	8
69 static struct hlist_head bsg_device_list[BSG_LIST_ARRAY_SIZE];
70 
71 static struct class *bsg_class;
72 static int bsg_major;
73 
74 static struct kmem_cache *bsg_cmd_cachep;
75 
76 /*
77  * our internal command type
78  */
79 struct bsg_command {
80 	struct bsg_device *bd;
81 	struct list_head list;
82 	struct request *rq;
83 	struct bio *bio;
84 	struct bio *bidi_bio;
85 	int err;
86 	struct sg_io_v4 hdr;
87 	char sense[SCSI_SENSE_BUFFERSIZE];
88 };
89 
90 static void bsg_free_command(struct bsg_command *bc)
91 {
92 	struct bsg_device *bd = bc->bd;
93 	unsigned long flags;
94 
95 	kmem_cache_free(bsg_cmd_cachep, bc);
96 
97 	spin_lock_irqsave(&bd->lock, flags);
98 	bd->queued_cmds--;
99 	spin_unlock_irqrestore(&bd->lock, flags);
100 
101 	wake_up(&bd->wq_free);
102 }
103 
104 static struct bsg_command *bsg_alloc_command(struct bsg_device *bd)
105 {
106 	struct bsg_command *bc = ERR_PTR(-EINVAL);
107 
108 	spin_lock_irq(&bd->lock);
109 
110 	if (bd->queued_cmds >= bd->max_queue)
111 		goto out;
112 
113 	bd->queued_cmds++;
114 	spin_unlock_irq(&bd->lock);
115 
116 	bc = kmem_cache_zalloc(bsg_cmd_cachep, GFP_KERNEL);
117 	if (unlikely(!bc)) {
118 		spin_lock_irq(&bd->lock);
119 		bd->queued_cmds--;
120 		bc = ERR_PTR(-ENOMEM);
121 		goto out;
122 	}
123 
124 	bc->bd = bd;
125 	INIT_LIST_HEAD(&bc->list);
126 	dprintk("%s: returning free cmd %p\n", bd->name, bc);
127 	return bc;
128 out:
129 	spin_unlock_irq(&bd->lock);
130 	return bc;
131 }
132 
133 static inline struct hlist_head *bsg_dev_idx_hash(int index)
134 {
135 	return &bsg_device_list[index & (BSG_LIST_ARRAY_SIZE - 1)];
136 }
137 
138 static int bsg_io_schedule(struct bsg_device *bd)
139 {
140 	DEFINE_WAIT(wait);
141 	int ret = 0;
142 
143 	spin_lock_irq(&bd->lock);
144 
145 	BUG_ON(bd->done_cmds > bd->queued_cmds);
146 
147 	/*
148 	 * -ENOSPC or -ENODATA?  I'm going for -ENODATA, meaning "I have no
149 	 * work to do", even though we return -ENOSPC after this same test
150 	 * during bsg_write() -- there, it means our buffer can't have more
151 	 * bsg_commands added to it, thus has no space left.
152 	 */
153 	if (bd->done_cmds == bd->queued_cmds) {
154 		ret = -ENODATA;
155 		goto unlock;
156 	}
157 
158 	if (!test_bit(BSG_F_BLOCK, &bd->flags)) {
159 		ret = -EAGAIN;
160 		goto unlock;
161 	}
162 
163 	prepare_to_wait(&bd->wq_done, &wait, TASK_UNINTERRUPTIBLE);
164 	spin_unlock_irq(&bd->lock);
165 	io_schedule();
166 	finish_wait(&bd->wq_done, &wait);
167 
168 	return ret;
169 unlock:
170 	spin_unlock_irq(&bd->lock);
171 	return ret;
172 }
173 
174 static int blk_fill_sgv4_hdr_rq(struct request_queue *q, struct request *rq,
175 				struct sg_io_v4 *hdr, struct bsg_device *bd,
176 				fmode_t has_write_perm)
177 {
178 	if (hdr->request_len > BLK_MAX_CDB) {
179 		rq->cmd = kzalloc(hdr->request_len, GFP_KERNEL);
180 		if (!rq->cmd)
181 			return -ENOMEM;
182 	}
183 
184 	if (copy_from_user(rq->cmd, (void *)(unsigned long)hdr->request,
185 			   hdr->request_len))
186 		return -EFAULT;
187 
188 	if (hdr->subprotocol == BSG_SUB_PROTOCOL_SCSI_CMD) {
189 		if (blk_verify_command(rq->cmd, has_write_perm))
190 			return -EPERM;
191 	} else if (!capable(CAP_SYS_RAWIO))
192 		return -EPERM;
193 
194 	/*
195 	 * fill in request structure
196 	 */
197 	rq->cmd_len = hdr->request_len;
198 	rq->cmd_type = REQ_TYPE_BLOCK_PC;
199 
200 	rq->timeout = (hdr->timeout * HZ) / 1000;
201 	if (!rq->timeout)
202 		rq->timeout = q->sg_timeout;
203 	if (!rq->timeout)
204 		rq->timeout = BLK_DEFAULT_SG_TIMEOUT;
205 	if (rq->timeout < BLK_MIN_SG_TIMEOUT)
206 		rq->timeout = BLK_MIN_SG_TIMEOUT;
207 
208 	return 0;
209 }
210 
211 /*
212  * Check if sg_io_v4 from user is allowed and valid
213  */
214 static int
215 bsg_validate_sgv4_hdr(struct request_queue *q, struct sg_io_v4 *hdr, int *rw)
216 {
217 	int ret = 0;
218 
219 	if (hdr->guard != 'Q')
220 		return -EINVAL;
221 
222 	switch (hdr->protocol) {
223 	case BSG_PROTOCOL_SCSI:
224 		switch (hdr->subprotocol) {
225 		case BSG_SUB_PROTOCOL_SCSI_CMD:
226 		case BSG_SUB_PROTOCOL_SCSI_TRANSPORT:
227 			break;
228 		default:
229 			ret = -EINVAL;
230 		}
231 		break;
232 	default:
233 		ret = -EINVAL;
234 	}
235 
236 	*rw = hdr->dout_xfer_len ? WRITE : READ;
237 	return ret;
238 }
239 
240 /*
241  * map sg_io_v4 to a request.
242  */
243 static struct request *
244 bsg_map_hdr(struct bsg_device *bd, struct sg_io_v4 *hdr, fmode_t has_write_perm,
245 	    u8 *sense)
246 {
247 	struct request_queue *q = bd->queue;
248 	struct request *rq, *next_rq = NULL;
249 	int ret, rw;
250 	unsigned int dxfer_len;
251 	void *dxferp = NULL;
252 
253 	dprintk("map hdr %llx/%u %llx/%u\n", (unsigned long long) hdr->dout_xferp,
254 		hdr->dout_xfer_len, (unsigned long long) hdr->din_xferp,
255 		hdr->din_xfer_len);
256 
257 	ret = bsg_validate_sgv4_hdr(q, hdr, &rw);
258 	if (ret)
259 		return ERR_PTR(ret);
260 
261 	/*
262 	 * map scatter-gather elements seperately and string them to request
263 	 */
264 	rq = blk_get_request(q, rw, GFP_KERNEL);
265 	if (!rq)
266 		return ERR_PTR(-ENOMEM);
267 	ret = blk_fill_sgv4_hdr_rq(q, rq, hdr, bd, has_write_perm);
268 	if (ret)
269 		goto out;
270 
271 	if (rw == WRITE && hdr->din_xfer_len) {
272 		if (!test_bit(QUEUE_FLAG_BIDI, &q->queue_flags)) {
273 			ret = -EOPNOTSUPP;
274 			goto out;
275 		}
276 
277 		next_rq = blk_get_request(q, READ, GFP_KERNEL);
278 		if (!next_rq) {
279 			ret = -ENOMEM;
280 			goto out;
281 		}
282 		rq->next_rq = next_rq;
283 		next_rq->cmd_type = rq->cmd_type;
284 
285 		dxferp = (void*)(unsigned long)hdr->din_xferp;
286 		ret =  blk_rq_map_user(q, next_rq, NULL, dxferp,
287 				       hdr->din_xfer_len, GFP_KERNEL);
288 		if (ret)
289 			goto out;
290 	}
291 
292 	if (hdr->dout_xfer_len) {
293 		dxfer_len = hdr->dout_xfer_len;
294 		dxferp = (void*)(unsigned long)hdr->dout_xferp;
295 	} else if (hdr->din_xfer_len) {
296 		dxfer_len = hdr->din_xfer_len;
297 		dxferp = (void*)(unsigned long)hdr->din_xferp;
298 	} else
299 		dxfer_len = 0;
300 
301 	if (dxfer_len) {
302 		ret = blk_rq_map_user(q, rq, NULL, dxferp, dxfer_len,
303 				      GFP_KERNEL);
304 		if (ret)
305 			goto out;
306 	}
307 
308 	rq->sense = sense;
309 	rq->sense_len = 0;
310 
311 	return rq;
312 out:
313 	if (rq->cmd != rq->__cmd)
314 		kfree(rq->cmd);
315 	blk_put_request(rq);
316 	if (next_rq) {
317 		blk_rq_unmap_user(next_rq->bio);
318 		blk_put_request(next_rq);
319 	}
320 	return ERR_PTR(ret);
321 }
322 
323 /*
324  * async completion call-back from the block layer, when scsi/ide/whatever
325  * calls end_that_request_last() on a request
326  */
327 static void bsg_rq_end_io(struct request *rq, int uptodate)
328 {
329 	struct bsg_command *bc = rq->end_io_data;
330 	struct bsg_device *bd = bc->bd;
331 	unsigned long flags;
332 
333 	dprintk("%s: finished rq %p bc %p, bio %p stat %d\n",
334 		bd->name, rq, bc, bc->bio, uptodate);
335 
336 	bc->hdr.duration = jiffies_to_msecs(jiffies - bc->hdr.duration);
337 
338 	spin_lock_irqsave(&bd->lock, flags);
339 	list_move_tail(&bc->list, &bd->done_list);
340 	bd->done_cmds++;
341 	spin_unlock_irqrestore(&bd->lock, flags);
342 
343 	wake_up(&bd->wq_done);
344 }
345 
346 /*
347  * do final setup of a 'bc' and submit the matching 'rq' to the block
348  * layer for io
349  */
350 static void bsg_add_command(struct bsg_device *bd, struct request_queue *q,
351 			    struct bsg_command *bc, struct request *rq)
352 {
353 	int at_head = (0 == (bc->hdr.flags & BSG_FLAG_Q_AT_TAIL));
354 
355 	/*
356 	 * add bc command to busy queue and submit rq for io
357 	 */
358 	bc->rq = rq;
359 	bc->bio = rq->bio;
360 	if (rq->next_rq)
361 		bc->bidi_bio = rq->next_rq->bio;
362 	bc->hdr.duration = jiffies;
363 	spin_lock_irq(&bd->lock);
364 	list_add_tail(&bc->list, &bd->busy_list);
365 	spin_unlock_irq(&bd->lock);
366 
367 	dprintk("%s: queueing rq %p, bc %p\n", bd->name, rq, bc);
368 
369 	rq->end_io_data = bc;
370 	blk_execute_rq_nowait(q, NULL, rq, at_head, bsg_rq_end_io);
371 }
372 
373 static struct bsg_command *bsg_next_done_cmd(struct bsg_device *bd)
374 {
375 	struct bsg_command *bc = NULL;
376 
377 	spin_lock_irq(&bd->lock);
378 	if (bd->done_cmds) {
379 		bc = list_first_entry(&bd->done_list, struct bsg_command, list);
380 		list_del(&bc->list);
381 		bd->done_cmds--;
382 	}
383 	spin_unlock_irq(&bd->lock);
384 
385 	return bc;
386 }
387 
388 /*
389  * Get a finished command from the done list
390  */
391 static struct bsg_command *bsg_get_done_cmd(struct bsg_device *bd)
392 {
393 	struct bsg_command *bc;
394 	int ret;
395 
396 	do {
397 		bc = bsg_next_done_cmd(bd);
398 		if (bc)
399 			break;
400 
401 		if (!test_bit(BSG_F_BLOCK, &bd->flags)) {
402 			bc = ERR_PTR(-EAGAIN);
403 			break;
404 		}
405 
406 		ret = wait_event_interruptible(bd->wq_done, bd->done_cmds);
407 		if (ret) {
408 			bc = ERR_PTR(-ERESTARTSYS);
409 			break;
410 		}
411 	} while (1);
412 
413 	dprintk("%s: returning done %p\n", bd->name, bc);
414 
415 	return bc;
416 }
417 
418 static int blk_complete_sgv4_hdr_rq(struct request *rq, struct sg_io_v4 *hdr,
419 				    struct bio *bio, struct bio *bidi_bio)
420 {
421 	int ret = 0;
422 
423 	dprintk("rq %p bio %p 0x%x\n", rq, bio, rq->errors);
424 	/*
425 	 * fill in all the output members
426 	 */
427 	hdr->device_status = status_byte(rq->errors);
428 	hdr->transport_status = host_byte(rq->errors);
429 	hdr->driver_status = driver_byte(rq->errors);
430 	hdr->info = 0;
431 	if (hdr->device_status || hdr->transport_status || hdr->driver_status)
432 		hdr->info |= SG_INFO_CHECK;
433 	hdr->response_len = 0;
434 
435 	if (rq->sense_len && hdr->response) {
436 		int len = min_t(unsigned int, hdr->max_response_len,
437 					rq->sense_len);
438 
439 		ret = copy_to_user((void*)(unsigned long)hdr->response,
440 				   rq->sense, len);
441 		if (!ret)
442 			hdr->response_len = len;
443 		else
444 			ret = -EFAULT;
445 	}
446 
447 	if (rq->next_rq) {
448 		hdr->dout_resid = rq->resid_len;
449 		hdr->din_resid = rq->next_rq->resid_len;
450 		blk_rq_unmap_user(bidi_bio);
451 		blk_put_request(rq->next_rq);
452 	} else if (rq_data_dir(rq) == READ)
453 		hdr->din_resid = rq->resid_len;
454 	else
455 		hdr->dout_resid = rq->resid_len;
456 
457 	/*
458 	 * If the request generated a negative error number, return it
459 	 * (providing we aren't already returning an error); if it's
460 	 * just a protocol response (i.e. non negative), that gets
461 	 * processed above.
462 	 */
463 	if (!ret && rq->errors < 0)
464 		ret = rq->errors;
465 
466 	blk_rq_unmap_user(bio);
467 	if (rq->cmd != rq->__cmd)
468 		kfree(rq->cmd);
469 	blk_put_request(rq);
470 
471 	return ret;
472 }
473 
474 static int bsg_complete_all_commands(struct bsg_device *bd)
475 {
476 	struct bsg_command *bc;
477 	int ret, tret;
478 
479 	dprintk("%s: entered\n", bd->name);
480 
481 	/*
482 	 * wait for all commands to complete
483 	 */
484 	ret = 0;
485 	do {
486 		ret = bsg_io_schedule(bd);
487 		/*
488 		 * look for -ENODATA specifically -- we'll sometimes get
489 		 * -ERESTARTSYS when we've taken a signal, but we can't
490 		 * return until we're done freeing the queue, so ignore
491 		 * it.  The signal will get handled when we're done freeing
492 		 * the bsg_device.
493 		 */
494 	} while (ret != -ENODATA);
495 
496 	/*
497 	 * discard done commands
498 	 */
499 	ret = 0;
500 	do {
501 		spin_lock_irq(&bd->lock);
502 		if (!bd->queued_cmds) {
503 			spin_unlock_irq(&bd->lock);
504 			break;
505 		}
506 		spin_unlock_irq(&bd->lock);
507 
508 		bc = bsg_get_done_cmd(bd);
509 		if (IS_ERR(bc))
510 			break;
511 
512 		tret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
513 						bc->bidi_bio);
514 		if (!ret)
515 			ret = tret;
516 
517 		bsg_free_command(bc);
518 	} while (1);
519 
520 	return ret;
521 }
522 
523 static int
524 __bsg_read(char __user *buf, size_t count, struct bsg_device *bd,
525 	   const struct iovec *iov, ssize_t *bytes_read)
526 {
527 	struct bsg_command *bc;
528 	int nr_commands, ret;
529 
530 	if (count % sizeof(struct sg_io_v4))
531 		return -EINVAL;
532 
533 	ret = 0;
534 	nr_commands = count / sizeof(struct sg_io_v4);
535 	while (nr_commands) {
536 		bc = bsg_get_done_cmd(bd);
537 		if (IS_ERR(bc)) {
538 			ret = PTR_ERR(bc);
539 			break;
540 		}
541 
542 		/*
543 		 * this is the only case where we need to copy data back
544 		 * after completing the request. so do that here,
545 		 * bsg_complete_work() cannot do that for us
546 		 */
547 		ret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
548 					       bc->bidi_bio);
549 
550 		if (copy_to_user(buf, &bc->hdr, sizeof(bc->hdr)))
551 			ret = -EFAULT;
552 
553 		bsg_free_command(bc);
554 
555 		if (ret)
556 			break;
557 
558 		buf += sizeof(struct sg_io_v4);
559 		*bytes_read += sizeof(struct sg_io_v4);
560 		nr_commands--;
561 	}
562 
563 	return ret;
564 }
565 
566 static inline void bsg_set_block(struct bsg_device *bd, struct file *file)
567 {
568 	if (file->f_flags & O_NONBLOCK)
569 		clear_bit(BSG_F_BLOCK, &bd->flags);
570 	else
571 		set_bit(BSG_F_BLOCK, &bd->flags);
572 }
573 
574 /*
575  * Check if the error is a "real" error that we should return.
576  */
577 static inline int err_block_err(int ret)
578 {
579 	if (ret && ret != -ENOSPC && ret != -ENODATA && ret != -EAGAIN)
580 		return 1;
581 
582 	return 0;
583 }
584 
585 static ssize_t
586 bsg_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
587 {
588 	struct bsg_device *bd = file->private_data;
589 	int ret;
590 	ssize_t bytes_read;
591 
592 	dprintk("%s: read %Zd bytes\n", bd->name, count);
593 
594 	bsg_set_block(bd, file);
595 
596 	bytes_read = 0;
597 	ret = __bsg_read(buf, count, bd, NULL, &bytes_read);
598 	*ppos = bytes_read;
599 
600 	if (!bytes_read || (bytes_read && err_block_err(ret)))
601 		bytes_read = ret;
602 
603 	return bytes_read;
604 }
605 
606 static int __bsg_write(struct bsg_device *bd, const char __user *buf,
607 		       size_t count, ssize_t *bytes_written,
608 		       fmode_t has_write_perm)
609 {
610 	struct bsg_command *bc;
611 	struct request *rq;
612 	int ret, nr_commands;
613 
614 	if (count % sizeof(struct sg_io_v4))
615 		return -EINVAL;
616 
617 	nr_commands = count / sizeof(struct sg_io_v4);
618 	rq = NULL;
619 	bc = NULL;
620 	ret = 0;
621 	while (nr_commands) {
622 		struct request_queue *q = bd->queue;
623 
624 		bc = bsg_alloc_command(bd);
625 		if (IS_ERR(bc)) {
626 			ret = PTR_ERR(bc);
627 			bc = NULL;
628 			break;
629 		}
630 
631 		if (copy_from_user(&bc->hdr, buf, sizeof(bc->hdr))) {
632 			ret = -EFAULT;
633 			break;
634 		}
635 
636 		/*
637 		 * get a request, fill in the blanks, and add to request queue
638 		 */
639 		rq = bsg_map_hdr(bd, &bc->hdr, has_write_perm, bc->sense);
640 		if (IS_ERR(rq)) {
641 			ret = PTR_ERR(rq);
642 			rq = NULL;
643 			break;
644 		}
645 
646 		bsg_add_command(bd, q, bc, rq);
647 		bc = NULL;
648 		rq = NULL;
649 		nr_commands--;
650 		buf += sizeof(struct sg_io_v4);
651 		*bytes_written += sizeof(struct sg_io_v4);
652 	}
653 
654 	if (bc)
655 		bsg_free_command(bc);
656 
657 	return ret;
658 }
659 
660 static ssize_t
661 bsg_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
662 {
663 	struct bsg_device *bd = file->private_data;
664 	ssize_t bytes_written;
665 	int ret;
666 
667 	dprintk("%s: write %Zd bytes\n", bd->name, count);
668 
669 	bsg_set_block(bd, file);
670 
671 	bytes_written = 0;
672 	ret = __bsg_write(bd, buf, count, &bytes_written,
673 			  file->f_mode & FMODE_WRITE);
674 
675 	*ppos = bytes_written;
676 
677 	/*
678 	 * return bytes written on non-fatal errors
679 	 */
680 	if (!bytes_written || (bytes_written && err_block_err(ret)))
681 		bytes_written = ret;
682 
683 	dprintk("%s: returning %Zd\n", bd->name, bytes_written);
684 	return bytes_written;
685 }
686 
687 static struct bsg_device *bsg_alloc_device(void)
688 {
689 	struct bsg_device *bd;
690 
691 	bd = kzalloc(sizeof(struct bsg_device), GFP_KERNEL);
692 	if (unlikely(!bd))
693 		return NULL;
694 
695 	spin_lock_init(&bd->lock);
696 
697 	bd->max_queue = BSG_DEFAULT_CMDS;
698 
699 	INIT_LIST_HEAD(&bd->busy_list);
700 	INIT_LIST_HEAD(&bd->done_list);
701 	INIT_HLIST_NODE(&bd->dev_list);
702 
703 	init_waitqueue_head(&bd->wq_free);
704 	init_waitqueue_head(&bd->wq_done);
705 	return bd;
706 }
707 
708 static void bsg_kref_release_function(struct kref *kref)
709 {
710 	struct bsg_class_device *bcd =
711 		container_of(kref, struct bsg_class_device, ref);
712 	struct device *parent = bcd->parent;
713 
714 	if (bcd->release)
715 		bcd->release(bcd->parent);
716 
717 	put_device(parent);
718 }
719 
720 static int bsg_put_device(struct bsg_device *bd)
721 {
722 	int ret = 0, do_free;
723 	struct request_queue *q = bd->queue;
724 
725 	mutex_lock(&bsg_mutex);
726 
727 	do_free = atomic_dec_and_test(&bd->ref_count);
728 	if (!do_free) {
729 		mutex_unlock(&bsg_mutex);
730 		goto out;
731 	}
732 
733 	hlist_del(&bd->dev_list);
734 	mutex_unlock(&bsg_mutex);
735 
736 	dprintk("%s: tearing down\n", bd->name);
737 
738 	/*
739 	 * close can always block
740 	 */
741 	set_bit(BSG_F_BLOCK, &bd->flags);
742 
743 	/*
744 	 * correct error detection baddies here again. it's the responsibility
745 	 * of the app to properly reap commands before close() if it wants
746 	 * fool-proof error detection
747 	 */
748 	ret = bsg_complete_all_commands(bd);
749 
750 	kfree(bd);
751 out:
752 	kref_put(&q->bsg_dev.ref, bsg_kref_release_function);
753 	if (do_free)
754 		blk_put_queue(q);
755 	return ret;
756 }
757 
758 static struct bsg_device *bsg_add_device(struct inode *inode,
759 					 struct request_queue *rq,
760 					 struct file *file)
761 {
762 	struct bsg_device *bd;
763 	int ret;
764 #ifdef BSG_DEBUG
765 	unsigned char buf[32];
766 #endif
767 	ret = blk_get_queue(rq);
768 	if (ret)
769 		return ERR_PTR(-ENXIO);
770 
771 	bd = bsg_alloc_device();
772 	if (!bd) {
773 		blk_put_queue(rq);
774 		return ERR_PTR(-ENOMEM);
775 	}
776 
777 	bd->queue = rq;
778 
779 	bsg_set_block(bd, file);
780 
781 	atomic_set(&bd->ref_count, 1);
782 	mutex_lock(&bsg_mutex);
783 	hlist_add_head(&bd->dev_list, bsg_dev_idx_hash(iminor(inode)));
784 
785 	strncpy(bd->name, dev_name(rq->bsg_dev.class_dev), sizeof(bd->name) - 1);
786 	dprintk("bound to <%s>, max queue %d\n",
787 		format_dev_t(buf, inode->i_rdev), bd->max_queue);
788 
789 	mutex_unlock(&bsg_mutex);
790 	return bd;
791 }
792 
793 static struct bsg_device *__bsg_get_device(int minor, struct request_queue *q)
794 {
795 	struct bsg_device *bd;
796 	struct hlist_node *entry;
797 
798 	mutex_lock(&bsg_mutex);
799 
800 	hlist_for_each_entry(bd, entry, bsg_dev_idx_hash(minor), dev_list) {
801 		if (bd->queue == q) {
802 			atomic_inc(&bd->ref_count);
803 			goto found;
804 		}
805 	}
806 	bd = NULL;
807 found:
808 	mutex_unlock(&bsg_mutex);
809 	return bd;
810 }
811 
812 static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file)
813 {
814 	struct bsg_device *bd;
815 	struct bsg_class_device *bcd;
816 
817 	/*
818 	 * find the class device
819 	 */
820 	mutex_lock(&bsg_mutex);
821 	bcd = idr_find(&bsg_minor_idr, iminor(inode));
822 	if (bcd)
823 		kref_get(&bcd->ref);
824 	mutex_unlock(&bsg_mutex);
825 
826 	if (!bcd)
827 		return ERR_PTR(-ENODEV);
828 
829 	bd = __bsg_get_device(iminor(inode), bcd->queue);
830 	if (bd)
831 		return bd;
832 
833 	bd = bsg_add_device(inode, bcd->queue, file);
834 	if (IS_ERR(bd))
835 		kref_put(&bcd->ref, bsg_kref_release_function);
836 
837 	return bd;
838 }
839 
840 static int bsg_open(struct inode *inode, struct file *file)
841 {
842 	struct bsg_device *bd;
843 
844 	lock_kernel();
845 	bd = bsg_get_device(inode, file);
846 	unlock_kernel();
847 
848 	if (IS_ERR(bd))
849 		return PTR_ERR(bd);
850 
851 	file->private_data = bd;
852 	return 0;
853 }
854 
855 static int bsg_release(struct inode *inode, struct file *file)
856 {
857 	struct bsg_device *bd = file->private_data;
858 
859 	file->private_data = NULL;
860 	return bsg_put_device(bd);
861 }
862 
863 static unsigned int bsg_poll(struct file *file, poll_table *wait)
864 {
865 	struct bsg_device *bd = file->private_data;
866 	unsigned int mask = 0;
867 
868 	poll_wait(file, &bd->wq_done, wait);
869 	poll_wait(file, &bd->wq_free, wait);
870 
871 	spin_lock_irq(&bd->lock);
872 	if (!list_empty(&bd->done_list))
873 		mask |= POLLIN | POLLRDNORM;
874 	if (bd->queued_cmds >= bd->max_queue)
875 		mask |= POLLOUT;
876 	spin_unlock_irq(&bd->lock);
877 
878 	return mask;
879 }
880 
881 static long bsg_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
882 {
883 	struct bsg_device *bd = file->private_data;
884 	int __user *uarg = (int __user *) arg;
885 	int ret;
886 
887 	switch (cmd) {
888 		/*
889 		 * our own ioctls
890 		 */
891 	case SG_GET_COMMAND_Q:
892 		return put_user(bd->max_queue, uarg);
893 	case SG_SET_COMMAND_Q: {
894 		int queue;
895 
896 		if (get_user(queue, uarg))
897 			return -EFAULT;
898 		if (queue < 1)
899 			return -EINVAL;
900 
901 		spin_lock_irq(&bd->lock);
902 		bd->max_queue = queue;
903 		spin_unlock_irq(&bd->lock);
904 		return 0;
905 	}
906 
907 	/*
908 	 * SCSI/sg ioctls
909 	 */
910 	case SG_GET_VERSION_NUM:
911 	case SCSI_IOCTL_GET_IDLUN:
912 	case SCSI_IOCTL_GET_BUS_NUMBER:
913 	case SG_SET_TIMEOUT:
914 	case SG_GET_TIMEOUT:
915 	case SG_GET_RESERVED_SIZE:
916 	case SG_SET_RESERVED_SIZE:
917 	case SG_EMULATED_HOST:
918 	case SCSI_IOCTL_SEND_COMMAND: {
919 		void __user *uarg = (void __user *) arg;
920 		return scsi_cmd_ioctl(bd->queue, NULL, file->f_mode, cmd, uarg);
921 	}
922 	case SG_IO: {
923 		struct request *rq;
924 		struct bio *bio, *bidi_bio = NULL;
925 		struct sg_io_v4 hdr;
926 		int at_head;
927 		u8 sense[SCSI_SENSE_BUFFERSIZE];
928 
929 		if (copy_from_user(&hdr, uarg, sizeof(hdr)))
930 			return -EFAULT;
931 
932 		rq = bsg_map_hdr(bd, &hdr, file->f_mode & FMODE_WRITE, sense);
933 		if (IS_ERR(rq))
934 			return PTR_ERR(rq);
935 
936 		bio = rq->bio;
937 		if (rq->next_rq)
938 			bidi_bio = rq->next_rq->bio;
939 
940 		at_head = (0 == (hdr.flags & BSG_FLAG_Q_AT_TAIL));
941 		blk_execute_rq(bd->queue, NULL, rq, at_head);
942 		ret = blk_complete_sgv4_hdr_rq(rq, &hdr, bio, bidi_bio);
943 
944 		if (copy_to_user(uarg, &hdr, sizeof(hdr)))
945 			return -EFAULT;
946 
947 		return ret;
948 	}
949 	/*
950 	 * block device ioctls
951 	 */
952 	default:
953 #if 0
954 		return ioctl_by_bdev(bd->bdev, cmd, arg);
955 #else
956 		return -ENOTTY;
957 #endif
958 	}
959 }
960 
961 static const struct file_operations bsg_fops = {
962 	.read		=	bsg_read,
963 	.write		=	bsg_write,
964 	.poll		=	bsg_poll,
965 	.open		=	bsg_open,
966 	.release	=	bsg_release,
967 	.unlocked_ioctl	=	bsg_ioctl,
968 	.owner		=	THIS_MODULE,
969 };
970 
971 void bsg_unregister_queue(struct request_queue *q)
972 {
973 	struct bsg_class_device *bcd = &q->bsg_dev;
974 
975 	if (!bcd->class_dev)
976 		return;
977 
978 	mutex_lock(&bsg_mutex);
979 	idr_remove(&bsg_minor_idr, bcd->minor);
980 	sysfs_remove_link(&q->kobj, "bsg");
981 	device_unregister(bcd->class_dev);
982 	bcd->class_dev = NULL;
983 	kref_put(&bcd->ref, bsg_kref_release_function);
984 	mutex_unlock(&bsg_mutex);
985 }
986 EXPORT_SYMBOL_GPL(bsg_unregister_queue);
987 
988 int bsg_register_queue(struct request_queue *q, struct device *parent,
989 		       const char *name, void (*release)(struct device *))
990 {
991 	struct bsg_class_device *bcd;
992 	dev_t dev;
993 	int ret, minor;
994 	struct device *class_dev = NULL;
995 	const char *devname;
996 
997 	if (name)
998 		devname = name;
999 	else
1000 		devname = dev_name(parent);
1001 
1002 	/*
1003 	 * we need a proper transport to send commands, not a stacked device
1004 	 */
1005 	if (!q->request_fn)
1006 		return 0;
1007 
1008 	bcd = &q->bsg_dev;
1009 	memset(bcd, 0, sizeof(*bcd));
1010 
1011 	mutex_lock(&bsg_mutex);
1012 
1013 	ret = idr_pre_get(&bsg_minor_idr, GFP_KERNEL);
1014 	if (!ret) {
1015 		ret = -ENOMEM;
1016 		goto unlock;
1017 	}
1018 
1019 	ret = idr_get_new(&bsg_minor_idr, bcd, &minor);
1020 	if (ret < 0)
1021 		goto unlock;
1022 
1023 	if (minor >= BSG_MAX_DEVS) {
1024 		printk(KERN_ERR "bsg: too many bsg devices\n");
1025 		ret = -EINVAL;
1026 		goto remove_idr;
1027 	}
1028 
1029 	bcd->minor = minor;
1030 	bcd->queue = q;
1031 	bcd->parent = get_device(parent);
1032 	bcd->release = release;
1033 	kref_init(&bcd->ref);
1034 	dev = MKDEV(bsg_major, bcd->minor);
1035 	class_dev = device_create(bsg_class, parent, dev, NULL, "%s", devname);
1036 	if (IS_ERR(class_dev)) {
1037 		ret = PTR_ERR(class_dev);
1038 		goto put_dev;
1039 	}
1040 	bcd->class_dev = class_dev;
1041 
1042 	if (q->kobj.sd) {
1043 		ret = sysfs_create_link(&q->kobj, &bcd->class_dev->kobj, "bsg");
1044 		if (ret)
1045 			goto unregister_class_dev;
1046 	}
1047 
1048 	mutex_unlock(&bsg_mutex);
1049 	return 0;
1050 
1051 unregister_class_dev:
1052 	device_unregister(class_dev);
1053 put_dev:
1054 	put_device(parent);
1055 remove_idr:
1056 	idr_remove(&bsg_minor_idr, minor);
1057 unlock:
1058 	mutex_unlock(&bsg_mutex);
1059 	return ret;
1060 }
1061 EXPORT_SYMBOL_GPL(bsg_register_queue);
1062 
1063 static struct cdev bsg_cdev;
1064 
1065 static char *bsg_nodename(struct device *dev)
1066 {
1067 	return kasprintf(GFP_KERNEL, "bsg/%s", dev_name(dev));
1068 }
1069 
1070 static int __init bsg_init(void)
1071 {
1072 	int ret, i;
1073 	dev_t devid;
1074 
1075 	bsg_cmd_cachep = kmem_cache_create("bsg_cmd",
1076 				sizeof(struct bsg_command), 0, 0, NULL);
1077 	if (!bsg_cmd_cachep) {
1078 		printk(KERN_ERR "bsg: failed creating slab cache\n");
1079 		return -ENOMEM;
1080 	}
1081 
1082 	for (i = 0; i < BSG_LIST_ARRAY_SIZE; i++)
1083 		INIT_HLIST_HEAD(&bsg_device_list[i]);
1084 
1085 	bsg_class = class_create(THIS_MODULE, "bsg");
1086 	if (IS_ERR(bsg_class)) {
1087 		ret = PTR_ERR(bsg_class);
1088 		goto destroy_kmemcache;
1089 	}
1090 	bsg_class->nodename = bsg_nodename;
1091 
1092 	ret = alloc_chrdev_region(&devid, 0, BSG_MAX_DEVS, "bsg");
1093 	if (ret)
1094 		goto destroy_bsg_class;
1095 
1096 	bsg_major = MAJOR(devid);
1097 
1098 	cdev_init(&bsg_cdev, &bsg_fops);
1099 	ret = cdev_add(&bsg_cdev, MKDEV(bsg_major, 0), BSG_MAX_DEVS);
1100 	if (ret)
1101 		goto unregister_chrdev;
1102 
1103 	printk(KERN_INFO BSG_DESCRIPTION " version " BSG_VERSION
1104 	       " loaded (major %d)\n", bsg_major);
1105 	return 0;
1106 unregister_chrdev:
1107 	unregister_chrdev_region(MKDEV(bsg_major, 0), BSG_MAX_DEVS);
1108 destroy_bsg_class:
1109 	class_destroy(bsg_class);
1110 destroy_kmemcache:
1111 	kmem_cache_destroy(bsg_cmd_cachep);
1112 	return ret;
1113 }
1114 
1115 MODULE_AUTHOR("Jens Axboe");
1116 MODULE_DESCRIPTION(BSG_DESCRIPTION);
1117 MODULE_LICENSE("GPL");
1118 
1119 device_initcall(bsg_init);
1120