xref: /linux/block/bsg.c (revision 93d90ad708b8da6efc0e487b66111aa9db7f70c7)
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/jiffies.h>
19 #include <linux/percpu.h>
20 #include <linux/uio.h>
21 #include <linux/idr.h>
22 #include <linux/bsg.h>
23 #include <linux/slab.h>
24 
25 #include <scsi/scsi.h>
26 #include <scsi/scsi_ioctl.h>
27 #include <scsi/scsi_cmnd.h>
28 #include <scsi/scsi_device.h>
29 #include <scsi/scsi_driver.h>
30 #include <scsi/sg.h>
31 
32 #define BSG_DESCRIPTION	"Block layer SCSI generic (bsg) driver"
33 #define BSG_VERSION	"0.4"
34 
35 struct bsg_device {
36 	struct request_queue *queue;
37 	spinlock_t lock;
38 	struct list_head busy_list;
39 	struct list_head done_list;
40 	struct hlist_node dev_list;
41 	atomic_t ref_count;
42 	int queued_cmds;
43 	int done_cmds;
44 	wait_queue_head_t wq_done;
45 	wait_queue_head_t wq_free;
46 	char name[20];
47 	int max_queue;
48 	unsigned long flags;
49 };
50 
51 enum {
52 	BSG_F_BLOCK		= 1,
53 };
54 
55 #define BSG_DEFAULT_CMDS	64
56 #define BSG_MAX_DEVS		32768
57 
58 #undef BSG_DEBUG
59 
60 #ifdef BSG_DEBUG
61 #define dprintk(fmt, args...) printk(KERN_ERR "%s: " fmt, __func__, ##args)
62 #else
63 #define dprintk(fmt, args...)
64 #endif
65 
66 static DEFINE_MUTEX(bsg_mutex);
67 static DEFINE_IDR(bsg_minor_idr);
68 
69 #define BSG_LIST_ARRAY_SIZE	8
70 static struct hlist_head bsg_device_list[BSG_LIST_ARRAY_SIZE];
71 
72 static struct class *bsg_class;
73 static int bsg_major;
74 
75 static struct kmem_cache *bsg_cmd_cachep;
76 
77 /*
78  * our internal command type
79  */
80 struct bsg_command {
81 	struct bsg_device *bd;
82 	struct list_head list;
83 	struct request *rq;
84 	struct bio *bio;
85 	struct bio *bidi_bio;
86 	int err;
87 	struct sg_io_v4 hdr;
88 	char sense[SCSI_SENSE_BUFFERSIZE];
89 };
90 
91 static void bsg_free_command(struct bsg_command *bc)
92 {
93 	struct bsg_device *bd = bc->bd;
94 	unsigned long flags;
95 
96 	kmem_cache_free(bsg_cmd_cachep, bc);
97 
98 	spin_lock_irqsave(&bd->lock, flags);
99 	bd->queued_cmds--;
100 	spin_unlock_irqrestore(&bd->lock, flags);
101 
102 	wake_up(&bd->wq_free);
103 }
104 
105 static struct bsg_command *bsg_alloc_command(struct bsg_device *bd)
106 {
107 	struct bsg_command *bc = ERR_PTR(-EINVAL);
108 
109 	spin_lock_irq(&bd->lock);
110 
111 	if (bd->queued_cmds >= bd->max_queue)
112 		goto out;
113 
114 	bd->queued_cmds++;
115 	spin_unlock_irq(&bd->lock);
116 
117 	bc = kmem_cache_zalloc(bsg_cmd_cachep, GFP_KERNEL);
118 	if (unlikely(!bc)) {
119 		spin_lock_irq(&bd->lock);
120 		bd->queued_cmds--;
121 		bc = ERR_PTR(-ENOMEM);
122 		goto out;
123 	}
124 
125 	bc->bd = bd;
126 	INIT_LIST_HEAD(&bc->list);
127 	dprintk("%s: returning free cmd %p\n", bd->name, bc);
128 	return bc;
129 out:
130 	spin_unlock_irq(&bd->lock);
131 	return bc;
132 }
133 
134 static inline struct hlist_head *bsg_dev_idx_hash(int index)
135 {
136 	return &bsg_device_list[index & (BSG_LIST_ARRAY_SIZE - 1)];
137 }
138 
139 static int bsg_io_schedule(struct bsg_device *bd)
140 {
141 	DEFINE_WAIT(wait);
142 	int ret = 0;
143 
144 	spin_lock_irq(&bd->lock);
145 
146 	BUG_ON(bd->done_cmds > bd->queued_cmds);
147 
148 	/*
149 	 * -ENOSPC or -ENODATA?  I'm going for -ENODATA, meaning "I have no
150 	 * work to do", even though we return -ENOSPC after this same test
151 	 * during bsg_write() -- there, it means our buffer can't have more
152 	 * bsg_commands added to it, thus has no space left.
153 	 */
154 	if (bd->done_cmds == bd->queued_cmds) {
155 		ret = -ENODATA;
156 		goto unlock;
157 	}
158 
159 	if (!test_bit(BSG_F_BLOCK, &bd->flags)) {
160 		ret = -EAGAIN;
161 		goto unlock;
162 	}
163 
164 	prepare_to_wait(&bd->wq_done, &wait, TASK_UNINTERRUPTIBLE);
165 	spin_unlock_irq(&bd->lock);
166 	io_schedule();
167 	finish_wait(&bd->wq_done, &wait);
168 
169 	return ret;
170 unlock:
171 	spin_unlock_irq(&bd->lock);
172 	return ret;
173 }
174 
175 static int blk_fill_sgv4_hdr_rq(struct request_queue *q, struct request *rq,
176 				struct sg_io_v4 *hdr, struct bsg_device *bd,
177 				fmode_t has_write_perm)
178 {
179 	if (hdr->request_len > BLK_MAX_CDB) {
180 		rq->cmd = kzalloc(hdr->request_len, GFP_KERNEL);
181 		if (!rq->cmd)
182 			return -ENOMEM;
183 	}
184 
185 	if (copy_from_user(rq->cmd, (void __user *)(unsigned long)hdr->request,
186 			   hdr->request_len))
187 		return -EFAULT;
188 
189 	if (hdr->subprotocol == BSG_SUB_PROTOCOL_SCSI_CMD) {
190 		if (blk_verify_command(rq->cmd, has_write_perm))
191 			return -EPERM;
192 	} else if (!capable(CAP_SYS_RAWIO))
193 		return -EPERM;
194 
195 	/*
196 	 * fill in request structure
197 	 */
198 	rq->cmd_len = hdr->request_len;
199 
200 	rq->timeout = msecs_to_jiffies(hdr->timeout);
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 __user *dxferp = NULL;
252 	struct bsg_class_device *bcd = &q->bsg_dev;
253 
254 	/* if the LLD has been removed then the bsg_unregister_queue will
255 	 * eventually be called and the class_dev was freed, so we can no
256 	 * longer use this request_queue. Return no such address.
257 	 */
258 	if (!bcd->class_dev)
259 		return ERR_PTR(-ENXIO);
260 
261 	dprintk("map hdr %llx/%u %llx/%u\n", (unsigned long long) hdr->dout_xferp,
262 		hdr->dout_xfer_len, (unsigned long long) hdr->din_xferp,
263 		hdr->din_xfer_len);
264 
265 	ret = bsg_validate_sgv4_hdr(q, hdr, &rw);
266 	if (ret)
267 		return ERR_PTR(ret);
268 
269 	/*
270 	 * map scatter-gather elements separately and string them to request
271 	 */
272 	rq = blk_get_request(q, rw, GFP_KERNEL);
273 	if (IS_ERR(rq))
274 		return rq;
275 	blk_rq_set_block_pc(rq);
276 
277 	ret = blk_fill_sgv4_hdr_rq(q, rq, hdr, bd, has_write_perm);
278 	if (ret)
279 		goto out;
280 
281 	if (rw == WRITE && hdr->din_xfer_len) {
282 		if (!test_bit(QUEUE_FLAG_BIDI, &q->queue_flags)) {
283 			ret = -EOPNOTSUPP;
284 			goto out;
285 		}
286 
287 		next_rq = blk_get_request(q, READ, GFP_KERNEL);
288 		if (IS_ERR(next_rq)) {
289 			ret = PTR_ERR(next_rq);
290 			next_rq = NULL;
291 			goto out;
292 		}
293 		rq->next_rq = next_rq;
294 		next_rq->cmd_type = rq->cmd_type;
295 
296 		dxferp = (void __user *)(unsigned long)hdr->din_xferp;
297 		ret =  blk_rq_map_user(q, next_rq, NULL, dxferp,
298 				       hdr->din_xfer_len, GFP_KERNEL);
299 		if (ret)
300 			goto out;
301 	}
302 
303 	if (hdr->dout_xfer_len) {
304 		dxfer_len = hdr->dout_xfer_len;
305 		dxferp = (void __user *)(unsigned long)hdr->dout_xferp;
306 	} else if (hdr->din_xfer_len) {
307 		dxfer_len = hdr->din_xfer_len;
308 		dxferp = (void __user *)(unsigned long)hdr->din_xferp;
309 	} else
310 		dxfer_len = 0;
311 
312 	if (dxfer_len) {
313 		ret = blk_rq_map_user(q, rq, NULL, dxferp, dxfer_len,
314 				      GFP_KERNEL);
315 		if (ret)
316 			goto out;
317 	}
318 
319 	rq->sense = sense;
320 	rq->sense_len = 0;
321 
322 	return rq;
323 out:
324 	if (rq->cmd != rq->__cmd)
325 		kfree(rq->cmd);
326 	blk_put_request(rq);
327 	if (next_rq) {
328 		blk_rq_unmap_user(next_rq->bio);
329 		blk_put_request(next_rq);
330 	}
331 	return ERR_PTR(ret);
332 }
333 
334 /*
335  * async completion call-back from the block layer, when scsi/ide/whatever
336  * calls end_that_request_last() on a request
337  */
338 static void bsg_rq_end_io(struct request *rq, int uptodate)
339 {
340 	struct bsg_command *bc = rq->end_io_data;
341 	struct bsg_device *bd = bc->bd;
342 	unsigned long flags;
343 
344 	dprintk("%s: finished rq %p bc %p, bio %p stat %d\n",
345 		bd->name, rq, bc, bc->bio, uptodate);
346 
347 	bc->hdr.duration = jiffies_to_msecs(jiffies - bc->hdr.duration);
348 
349 	spin_lock_irqsave(&bd->lock, flags);
350 	list_move_tail(&bc->list, &bd->done_list);
351 	bd->done_cmds++;
352 	spin_unlock_irqrestore(&bd->lock, flags);
353 
354 	wake_up(&bd->wq_done);
355 }
356 
357 /*
358  * do final setup of a 'bc' and submit the matching 'rq' to the block
359  * layer for io
360  */
361 static void bsg_add_command(struct bsg_device *bd, struct request_queue *q,
362 			    struct bsg_command *bc, struct request *rq)
363 {
364 	int at_head = (0 == (bc->hdr.flags & BSG_FLAG_Q_AT_TAIL));
365 
366 	/*
367 	 * add bc command to busy queue and submit rq for io
368 	 */
369 	bc->rq = rq;
370 	bc->bio = rq->bio;
371 	if (rq->next_rq)
372 		bc->bidi_bio = rq->next_rq->bio;
373 	bc->hdr.duration = jiffies;
374 	spin_lock_irq(&bd->lock);
375 	list_add_tail(&bc->list, &bd->busy_list);
376 	spin_unlock_irq(&bd->lock);
377 
378 	dprintk("%s: queueing rq %p, bc %p\n", bd->name, rq, bc);
379 
380 	rq->end_io_data = bc;
381 	blk_execute_rq_nowait(q, NULL, rq, at_head, bsg_rq_end_io);
382 }
383 
384 static struct bsg_command *bsg_next_done_cmd(struct bsg_device *bd)
385 {
386 	struct bsg_command *bc = NULL;
387 
388 	spin_lock_irq(&bd->lock);
389 	if (bd->done_cmds) {
390 		bc = list_first_entry(&bd->done_list, struct bsg_command, list);
391 		list_del(&bc->list);
392 		bd->done_cmds--;
393 	}
394 	spin_unlock_irq(&bd->lock);
395 
396 	return bc;
397 }
398 
399 /*
400  * Get a finished command from the done list
401  */
402 static struct bsg_command *bsg_get_done_cmd(struct bsg_device *bd)
403 {
404 	struct bsg_command *bc;
405 	int ret;
406 
407 	do {
408 		bc = bsg_next_done_cmd(bd);
409 		if (bc)
410 			break;
411 
412 		if (!test_bit(BSG_F_BLOCK, &bd->flags)) {
413 			bc = ERR_PTR(-EAGAIN);
414 			break;
415 		}
416 
417 		ret = wait_event_interruptible(bd->wq_done, bd->done_cmds);
418 		if (ret) {
419 			bc = ERR_PTR(-ERESTARTSYS);
420 			break;
421 		}
422 	} while (1);
423 
424 	dprintk("%s: returning done %p\n", bd->name, bc);
425 
426 	return bc;
427 }
428 
429 static int blk_complete_sgv4_hdr_rq(struct request *rq, struct sg_io_v4 *hdr,
430 				    struct bio *bio, struct bio *bidi_bio)
431 {
432 	int ret = 0;
433 
434 	dprintk("rq %p bio %p 0x%x\n", rq, bio, rq->errors);
435 	/*
436 	 * fill in all the output members
437 	 */
438 	hdr->device_status = rq->errors & 0xff;
439 	hdr->transport_status = host_byte(rq->errors);
440 	hdr->driver_status = driver_byte(rq->errors);
441 	hdr->info = 0;
442 	if (hdr->device_status || hdr->transport_status || hdr->driver_status)
443 		hdr->info |= SG_INFO_CHECK;
444 	hdr->response_len = 0;
445 
446 	if (rq->sense_len && hdr->response) {
447 		int len = min_t(unsigned int, hdr->max_response_len,
448 					rq->sense_len);
449 
450 		ret = copy_to_user((void __user *)(unsigned long)hdr->response,
451 				   rq->sense, len);
452 		if (!ret)
453 			hdr->response_len = len;
454 		else
455 			ret = -EFAULT;
456 	}
457 
458 	if (rq->next_rq) {
459 		hdr->dout_resid = rq->resid_len;
460 		hdr->din_resid = rq->next_rq->resid_len;
461 		blk_rq_unmap_user(bidi_bio);
462 		blk_put_request(rq->next_rq);
463 	} else if (rq_data_dir(rq) == READ)
464 		hdr->din_resid = rq->resid_len;
465 	else
466 		hdr->dout_resid = rq->resid_len;
467 
468 	/*
469 	 * If the request generated a negative error number, return it
470 	 * (providing we aren't already returning an error); if it's
471 	 * just a protocol response (i.e. non negative), that gets
472 	 * processed above.
473 	 */
474 	if (!ret && rq->errors < 0)
475 		ret = rq->errors;
476 
477 	blk_rq_unmap_user(bio);
478 	if (rq->cmd != rq->__cmd)
479 		kfree(rq->cmd);
480 	blk_put_request(rq);
481 
482 	return ret;
483 }
484 
485 static int bsg_complete_all_commands(struct bsg_device *bd)
486 {
487 	struct bsg_command *bc;
488 	int ret, tret;
489 
490 	dprintk("%s: entered\n", bd->name);
491 
492 	/*
493 	 * wait for all commands to complete
494 	 */
495 	ret = 0;
496 	do {
497 		ret = bsg_io_schedule(bd);
498 		/*
499 		 * look for -ENODATA specifically -- we'll sometimes get
500 		 * -ERESTARTSYS when we've taken a signal, but we can't
501 		 * return until we're done freeing the queue, so ignore
502 		 * it.  The signal will get handled when we're done freeing
503 		 * the bsg_device.
504 		 */
505 	} while (ret != -ENODATA);
506 
507 	/*
508 	 * discard done commands
509 	 */
510 	ret = 0;
511 	do {
512 		spin_lock_irq(&bd->lock);
513 		if (!bd->queued_cmds) {
514 			spin_unlock_irq(&bd->lock);
515 			break;
516 		}
517 		spin_unlock_irq(&bd->lock);
518 
519 		bc = bsg_get_done_cmd(bd);
520 		if (IS_ERR(bc))
521 			break;
522 
523 		tret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
524 						bc->bidi_bio);
525 		if (!ret)
526 			ret = tret;
527 
528 		bsg_free_command(bc);
529 	} while (1);
530 
531 	return ret;
532 }
533 
534 static int
535 __bsg_read(char __user *buf, size_t count, struct bsg_device *bd,
536 	   const struct iovec *iov, ssize_t *bytes_read)
537 {
538 	struct bsg_command *bc;
539 	int nr_commands, ret;
540 
541 	if (count % sizeof(struct sg_io_v4))
542 		return -EINVAL;
543 
544 	ret = 0;
545 	nr_commands = count / sizeof(struct sg_io_v4);
546 	while (nr_commands) {
547 		bc = bsg_get_done_cmd(bd);
548 		if (IS_ERR(bc)) {
549 			ret = PTR_ERR(bc);
550 			break;
551 		}
552 
553 		/*
554 		 * this is the only case where we need to copy data back
555 		 * after completing the request. so do that here,
556 		 * bsg_complete_work() cannot do that for us
557 		 */
558 		ret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
559 					       bc->bidi_bio);
560 
561 		if (copy_to_user(buf, &bc->hdr, sizeof(bc->hdr)))
562 			ret = -EFAULT;
563 
564 		bsg_free_command(bc);
565 
566 		if (ret)
567 			break;
568 
569 		buf += sizeof(struct sg_io_v4);
570 		*bytes_read += sizeof(struct sg_io_v4);
571 		nr_commands--;
572 	}
573 
574 	return ret;
575 }
576 
577 static inline void bsg_set_block(struct bsg_device *bd, struct file *file)
578 {
579 	if (file->f_flags & O_NONBLOCK)
580 		clear_bit(BSG_F_BLOCK, &bd->flags);
581 	else
582 		set_bit(BSG_F_BLOCK, &bd->flags);
583 }
584 
585 /*
586  * Check if the error is a "real" error that we should return.
587  */
588 static inline int err_block_err(int ret)
589 {
590 	if (ret && ret != -ENOSPC && ret != -ENODATA && ret != -EAGAIN)
591 		return 1;
592 
593 	return 0;
594 }
595 
596 static ssize_t
597 bsg_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
598 {
599 	struct bsg_device *bd = file->private_data;
600 	int ret;
601 	ssize_t bytes_read;
602 
603 	dprintk("%s: read %Zd bytes\n", bd->name, count);
604 
605 	bsg_set_block(bd, file);
606 
607 	bytes_read = 0;
608 	ret = __bsg_read(buf, count, bd, NULL, &bytes_read);
609 	*ppos = bytes_read;
610 
611 	if (!bytes_read || err_block_err(ret))
612 		bytes_read = ret;
613 
614 	return bytes_read;
615 }
616 
617 static int __bsg_write(struct bsg_device *bd, const char __user *buf,
618 		       size_t count, ssize_t *bytes_written,
619 		       fmode_t has_write_perm)
620 {
621 	struct bsg_command *bc;
622 	struct request *rq;
623 	int ret, nr_commands;
624 
625 	if (count % sizeof(struct sg_io_v4))
626 		return -EINVAL;
627 
628 	nr_commands = count / sizeof(struct sg_io_v4);
629 	rq = NULL;
630 	bc = NULL;
631 	ret = 0;
632 	while (nr_commands) {
633 		struct request_queue *q = bd->queue;
634 
635 		bc = bsg_alloc_command(bd);
636 		if (IS_ERR(bc)) {
637 			ret = PTR_ERR(bc);
638 			bc = NULL;
639 			break;
640 		}
641 
642 		if (copy_from_user(&bc->hdr, buf, sizeof(bc->hdr))) {
643 			ret = -EFAULT;
644 			break;
645 		}
646 
647 		/*
648 		 * get a request, fill in the blanks, and add to request queue
649 		 */
650 		rq = bsg_map_hdr(bd, &bc->hdr, has_write_perm, bc->sense);
651 		if (IS_ERR(rq)) {
652 			ret = PTR_ERR(rq);
653 			rq = NULL;
654 			break;
655 		}
656 
657 		bsg_add_command(bd, q, bc, rq);
658 		bc = NULL;
659 		rq = NULL;
660 		nr_commands--;
661 		buf += sizeof(struct sg_io_v4);
662 		*bytes_written += sizeof(struct sg_io_v4);
663 	}
664 
665 	if (bc)
666 		bsg_free_command(bc);
667 
668 	return ret;
669 }
670 
671 static ssize_t
672 bsg_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
673 {
674 	struct bsg_device *bd = file->private_data;
675 	ssize_t bytes_written;
676 	int ret;
677 
678 	dprintk("%s: write %Zd bytes\n", bd->name, count);
679 
680 	bsg_set_block(bd, file);
681 
682 	bytes_written = 0;
683 	ret = __bsg_write(bd, buf, count, &bytes_written,
684 			  file->f_mode & FMODE_WRITE);
685 
686 	*ppos = bytes_written;
687 
688 	/*
689 	 * return bytes written on non-fatal errors
690 	 */
691 	if (!bytes_written || err_block_err(ret))
692 		bytes_written = ret;
693 
694 	dprintk("%s: returning %Zd\n", bd->name, bytes_written);
695 	return bytes_written;
696 }
697 
698 static struct bsg_device *bsg_alloc_device(void)
699 {
700 	struct bsg_device *bd;
701 
702 	bd = kzalloc(sizeof(struct bsg_device), GFP_KERNEL);
703 	if (unlikely(!bd))
704 		return NULL;
705 
706 	spin_lock_init(&bd->lock);
707 
708 	bd->max_queue = BSG_DEFAULT_CMDS;
709 
710 	INIT_LIST_HEAD(&bd->busy_list);
711 	INIT_LIST_HEAD(&bd->done_list);
712 	INIT_HLIST_NODE(&bd->dev_list);
713 
714 	init_waitqueue_head(&bd->wq_free);
715 	init_waitqueue_head(&bd->wq_done);
716 	return bd;
717 }
718 
719 static void bsg_kref_release_function(struct kref *kref)
720 {
721 	struct bsg_class_device *bcd =
722 		container_of(kref, struct bsg_class_device, ref);
723 	struct device *parent = bcd->parent;
724 
725 	if (bcd->release)
726 		bcd->release(bcd->parent);
727 
728 	put_device(parent);
729 }
730 
731 static int bsg_put_device(struct bsg_device *bd)
732 {
733 	int ret = 0, do_free;
734 	struct request_queue *q = bd->queue;
735 
736 	mutex_lock(&bsg_mutex);
737 
738 	do_free = atomic_dec_and_test(&bd->ref_count);
739 	if (!do_free) {
740 		mutex_unlock(&bsg_mutex);
741 		goto out;
742 	}
743 
744 	hlist_del(&bd->dev_list);
745 	mutex_unlock(&bsg_mutex);
746 
747 	dprintk("%s: tearing down\n", bd->name);
748 
749 	/*
750 	 * close can always block
751 	 */
752 	set_bit(BSG_F_BLOCK, &bd->flags);
753 
754 	/*
755 	 * correct error detection baddies here again. it's the responsibility
756 	 * of the app to properly reap commands before close() if it wants
757 	 * fool-proof error detection
758 	 */
759 	ret = bsg_complete_all_commands(bd);
760 
761 	kfree(bd);
762 out:
763 	kref_put(&q->bsg_dev.ref, bsg_kref_release_function);
764 	if (do_free)
765 		blk_put_queue(q);
766 	return ret;
767 }
768 
769 static struct bsg_device *bsg_add_device(struct inode *inode,
770 					 struct request_queue *rq,
771 					 struct file *file)
772 {
773 	struct bsg_device *bd;
774 #ifdef BSG_DEBUG
775 	unsigned char buf[32];
776 #endif
777 	if (!blk_get_queue(rq))
778 		return ERR_PTR(-ENXIO);
779 
780 	bd = bsg_alloc_device();
781 	if (!bd) {
782 		blk_put_queue(rq);
783 		return ERR_PTR(-ENOMEM);
784 	}
785 
786 	bd->queue = rq;
787 
788 	bsg_set_block(bd, file);
789 
790 	atomic_set(&bd->ref_count, 1);
791 	mutex_lock(&bsg_mutex);
792 	hlist_add_head(&bd->dev_list, bsg_dev_idx_hash(iminor(inode)));
793 
794 	strncpy(bd->name, dev_name(rq->bsg_dev.class_dev), sizeof(bd->name) - 1);
795 	dprintk("bound to <%s>, max queue %d\n",
796 		format_dev_t(buf, inode->i_rdev), bd->max_queue);
797 
798 	mutex_unlock(&bsg_mutex);
799 	return bd;
800 }
801 
802 static struct bsg_device *__bsg_get_device(int minor, struct request_queue *q)
803 {
804 	struct bsg_device *bd;
805 
806 	mutex_lock(&bsg_mutex);
807 
808 	hlist_for_each_entry(bd, bsg_dev_idx_hash(minor), dev_list) {
809 		if (bd->queue == q) {
810 			atomic_inc(&bd->ref_count);
811 			goto found;
812 		}
813 	}
814 	bd = NULL;
815 found:
816 	mutex_unlock(&bsg_mutex);
817 	return bd;
818 }
819 
820 static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file)
821 {
822 	struct bsg_device *bd;
823 	struct bsg_class_device *bcd;
824 
825 	/*
826 	 * find the class device
827 	 */
828 	mutex_lock(&bsg_mutex);
829 	bcd = idr_find(&bsg_minor_idr, iminor(inode));
830 	if (bcd)
831 		kref_get(&bcd->ref);
832 	mutex_unlock(&bsg_mutex);
833 
834 	if (!bcd)
835 		return ERR_PTR(-ENODEV);
836 
837 	bd = __bsg_get_device(iminor(inode), bcd->queue);
838 	if (bd)
839 		return bd;
840 
841 	bd = bsg_add_device(inode, bcd->queue, file);
842 	if (IS_ERR(bd))
843 		kref_put(&bcd->ref, bsg_kref_release_function);
844 
845 	return bd;
846 }
847 
848 static int bsg_open(struct inode *inode, struct file *file)
849 {
850 	struct bsg_device *bd;
851 
852 	bd = bsg_get_device(inode, file);
853 
854 	if (IS_ERR(bd))
855 		return PTR_ERR(bd);
856 
857 	file->private_data = bd;
858 	return 0;
859 }
860 
861 static int bsg_release(struct inode *inode, struct file *file)
862 {
863 	struct bsg_device *bd = file->private_data;
864 
865 	file->private_data = NULL;
866 	return bsg_put_device(bd);
867 }
868 
869 static unsigned int bsg_poll(struct file *file, poll_table *wait)
870 {
871 	struct bsg_device *bd = file->private_data;
872 	unsigned int mask = 0;
873 
874 	poll_wait(file, &bd->wq_done, wait);
875 	poll_wait(file, &bd->wq_free, wait);
876 
877 	spin_lock_irq(&bd->lock);
878 	if (!list_empty(&bd->done_list))
879 		mask |= POLLIN | POLLRDNORM;
880 	if (bd->queued_cmds < bd->max_queue)
881 		mask |= POLLOUT;
882 	spin_unlock_irq(&bd->lock);
883 
884 	return mask;
885 }
886 
887 static long bsg_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
888 {
889 	struct bsg_device *bd = file->private_data;
890 	int __user *uarg = (int __user *) arg;
891 	int ret;
892 
893 	switch (cmd) {
894 		/*
895 		 * our own ioctls
896 		 */
897 	case SG_GET_COMMAND_Q:
898 		return put_user(bd->max_queue, uarg);
899 	case SG_SET_COMMAND_Q: {
900 		int queue;
901 
902 		if (get_user(queue, uarg))
903 			return -EFAULT;
904 		if (queue < 1)
905 			return -EINVAL;
906 
907 		spin_lock_irq(&bd->lock);
908 		bd->max_queue = queue;
909 		spin_unlock_irq(&bd->lock);
910 		return 0;
911 	}
912 
913 	/*
914 	 * SCSI/sg ioctls
915 	 */
916 	case SG_GET_VERSION_NUM:
917 	case SCSI_IOCTL_GET_IDLUN:
918 	case SCSI_IOCTL_GET_BUS_NUMBER:
919 	case SG_SET_TIMEOUT:
920 	case SG_GET_TIMEOUT:
921 	case SG_GET_RESERVED_SIZE:
922 	case SG_SET_RESERVED_SIZE:
923 	case SG_EMULATED_HOST:
924 	case SCSI_IOCTL_SEND_COMMAND: {
925 		void __user *uarg = (void __user *) arg;
926 		return scsi_cmd_ioctl(bd->queue, NULL, file->f_mode, cmd, uarg);
927 	}
928 	case SG_IO: {
929 		struct request *rq;
930 		struct bio *bio, *bidi_bio = NULL;
931 		struct sg_io_v4 hdr;
932 		int at_head;
933 		u8 sense[SCSI_SENSE_BUFFERSIZE];
934 
935 		if (copy_from_user(&hdr, uarg, sizeof(hdr)))
936 			return -EFAULT;
937 
938 		rq = bsg_map_hdr(bd, &hdr, file->f_mode & FMODE_WRITE, sense);
939 		if (IS_ERR(rq))
940 			return PTR_ERR(rq);
941 
942 		bio = rq->bio;
943 		if (rq->next_rq)
944 			bidi_bio = rq->next_rq->bio;
945 
946 		at_head = (0 == (hdr.flags & BSG_FLAG_Q_AT_TAIL));
947 		blk_execute_rq(bd->queue, NULL, rq, at_head);
948 		ret = blk_complete_sgv4_hdr_rq(rq, &hdr, bio, bidi_bio);
949 
950 		if (copy_to_user(uarg, &hdr, sizeof(hdr)))
951 			return -EFAULT;
952 
953 		return ret;
954 	}
955 	/*
956 	 * block device ioctls
957 	 */
958 	default:
959 #if 0
960 		return ioctl_by_bdev(bd->bdev, cmd, arg);
961 #else
962 		return -ENOTTY;
963 #endif
964 	}
965 }
966 
967 static const struct file_operations bsg_fops = {
968 	.read		=	bsg_read,
969 	.write		=	bsg_write,
970 	.poll		=	bsg_poll,
971 	.open		=	bsg_open,
972 	.release	=	bsg_release,
973 	.unlocked_ioctl	=	bsg_ioctl,
974 	.owner		=	THIS_MODULE,
975 	.llseek		=	default_llseek,
976 };
977 
978 void bsg_unregister_queue(struct request_queue *q)
979 {
980 	struct bsg_class_device *bcd = &q->bsg_dev;
981 
982 	if (!bcd->class_dev)
983 		return;
984 
985 	mutex_lock(&bsg_mutex);
986 	idr_remove(&bsg_minor_idr, bcd->minor);
987 	if (q->kobj.sd)
988 		sysfs_remove_link(&q->kobj, "bsg");
989 	device_unregister(bcd->class_dev);
990 	bcd->class_dev = NULL;
991 	kref_put(&bcd->ref, bsg_kref_release_function);
992 	mutex_unlock(&bsg_mutex);
993 }
994 EXPORT_SYMBOL_GPL(bsg_unregister_queue);
995 
996 int bsg_register_queue(struct request_queue *q, struct device *parent,
997 		       const char *name, void (*release)(struct device *))
998 {
999 	struct bsg_class_device *bcd;
1000 	dev_t dev;
1001 	int ret;
1002 	struct device *class_dev = NULL;
1003 	const char *devname;
1004 
1005 	if (name)
1006 		devname = name;
1007 	else
1008 		devname = dev_name(parent);
1009 
1010 	/*
1011 	 * we need a proper transport to send commands, not a stacked device
1012 	 */
1013 	if (!queue_is_rq_based(q))
1014 		return 0;
1015 
1016 	bcd = &q->bsg_dev;
1017 	memset(bcd, 0, sizeof(*bcd));
1018 
1019 	mutex_lock(&bsg_mutex);
1020 
1021 	ret = idr_alloc(&bsg_minor_idr, bcd, 0, BSG_MAX_DEVS, GFP_KERNEL);
1022 	if (ret < 0) {
1023 		if (ret == -ENOSPC) {
1024 			printk(KERN_ERR "bsg: too many bsg devices\n");
1025 			ret = -EINVAL;
1026 		}
1027 		goto unlock;
1028 	}
1029 
1030 	bcd->minor = ret;
1031 	bcd->queue = q;
1032 	bcd->parent = get_device(parent);
1033 	bcd->release = release;
1034 	kref_init(&bcd->ref);
1035 	dev = MKDEV(bsg_major, bcd->minor);
1036 	class_dev = device_create(bsg_class, parent, dev, NULL, "%s", devname);
1037 	if (IS_ERR(class_dev)) {
1038 		ret = PTR_ERR(class_dev);
1039 		goto put_dev;
1040 	}
1041 	bcd->class_dev = class_dev;
1042 
1043 	if (q->kobj.sd) {
1044 		ret = sysfs_create_link(&q->kobj, &bcd->class_dev->kobj, "bsg");
1045 		if (ret)
1046 			goto unregister_class_dev;
1047 	}
1048 
1049 	mutex_unlock(&bsg_mutex);
1050 	return 0;
1051 
1052 unregister_class_dev:
1053 	device_unregister(class_dev);
1054 put_dev:
1055 	put_device(parent);
1056 	idr_remove(&bsg_minor_idr, bcd->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_devnode(struct device *dev, umode_t *mode)
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->devnode = bsg_devnode;
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