xref: /openbmc/linux/block/bsg.c (revision f30828a6)
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[BUS_ID_SIZE];
46 	int max_queue;
47 	unsigned long flags;
48 	struct blk_scsi_cmd_filter *cmd_filter;
49 	mode_t *f_mode;
50 };
51 
52 enum {
53 	BSG_F_BLOCK		= 1,
54 };
55 
56 #define BSG_DEFAULT_CMDS	64
57 #define BSG_MAX_DEVS		32768
58 
59 #undef BSG_DEBUG
60 
61 #ifdef BSG_DEBUG
62 #define dprintk(fmt, args...) printk(KERN_ERR "%s: " fmt, __func__, ##args)
63 #else
64 #define dprintk(fmt, args...)
65 #endif
66 
67 static DEFINE_MUTEX(bsg_mutex);
68 static DEFINE_IDR(bsg_minor_idr);
69 
70 #define BSG_LIST_ARRAY_SIZE	8
71 static struct hlist_head bsg_device_list[BSG_LIST_ARRAY_SIZE];
72 
73 static struct class *bsg_class;
74 static int bsg_major;
75 
76 static struct kmem_cache *bsg_cmd_cachep;
77 
78 /*
79  * our internal command type
80  */
81 struct bsg_command {
82 	struct bsg_device *bd;
83 	struct list_head list;
84 	struct request *rq;
85 	struct bio *bio;
86 	struct bio *bidi_bio;
87 	int err;
88 	struct sg_io_v4 hdr;
89 	char sense[SCSI_SENSE_BUFFERSIZE];
90 };
91 
92 static void bsg_free_command(struct bsg_command *bc)
93 {
94 	struct bsg_device *bd = bc->bd;
95 	unsigned long flags;
96 
97 	kmem_cache_free(bsg_cmd_cachep, bc);
98 
99 	spin_lock_irqsave(&bd->lock, flags);
100 	bd->queued_cmds--;
101 	spin_unlock_irqrestore(&bd->lock, flags);
102 
103 	wake_up(&bd->wq_free);
104 }
105 
106 static struct bsg_command *bsg_alloc_command(struct bsg_device *bd)
107 {
108 	struct bsg_command *bc = ERR_PTR(-EINVAL);
109 
110 	spin_lock_irq(&bd->lock);
111 
112 	if (bd->queued_cmds >= bd->max_queue)
113 		goto out;
114 
115 	bd->queued_cmds++;
116 	spin_unlock_irq(&bd->lock);
117 
118 	bc = kmem_cache_zalloc(bsg_cmd_cachep, GFP_KERNEL);
119 	if (unlikely(!bc)) {
120 		spin_lock_irq(&bd->lock);
121 		bd->queued_cmds--;
122 		bc = ERR_PTR(-ENOMEM);
123 		goto out;
124 	}
125 
126 	bc->bd = bd;
127 	INIT_LIST_HEAD(&bc->list);
128 	dprintk("%s: returning free cmd %p\n", bd->name, bc);
129 	return bc;
130 out:
131 	spin_unlock_irq(&bd->lock);
132 	return bc;
133 }
134 
135 static inline struct hlist_head *bsg_dev_idx_hash(int index)
136 {
137 	return &bsg_device_list[index & (BSG_LIST_ARRAY_SIZE - 1)];
138 }
139 
140 static int bsg_io_schedule(struct bsg_device *bd)
141 {
142 	DEFINE_WAIT(wait);
143 	int ret = 0;
144 
145 	spin_lock_irq(&bd->lock);
146 
147 	BUG_ON(bd->done_cmds > bd->queued_cmds);
148 
149 	/*
150 	 * -ENOSPC or -ENODATA?  I'm going for -ENODATA, meaning "I have no
151 	 * work to do", even though we return -ENOSPC after this same test
152 	 * during bsg_write() -- there, it means our buffer can't have more
153 	 * bsg_commands added to it, thus has no space left.
154 	 */
155 	if (bd->done_cmds == bd->queued_cmds) {
156 		ret = -ENODATA;
157 		goto unlock;
158 	}
159 
160 	if (!test_bit(BSG_F_BLOCK, &bd->flags)) {
161 		ret = -EAGAIN;
162 		goto unlock;
163 	}
164 
165 	prepare_to_wait(&bd->wq_done, &wait, TASK_UNINTERRUPTIBLE);
166 	spin_unlock_irq(&bd->lock);
167 	io_schedule();
168 	finish_wait(&bd->wq_done, &wait);
169 
170 	return ret;
171 unlock:
172 	spin_unlock_irq(&bd->lock);
173 	return ret;
174 }
175 
176 static int blk_fill_sgv4_hdr_rq(struct request_queue *q, struct request *rq,
177 				struct sg_io_v4 *hdr, struct bsg_device *bd)
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 *)(unsigned long)hdr->request,
186 			   hdr->request_len))
187 		return -EFAULT;
188 
189 	if (hdr->subprotocol == BSG_SUB_PROTOCOL_SCSI_CMD) {
190 		if (blk_cmd_filter_verify_command(bd->cmd_filter, rq->cmd,
191 						 bd->f_mode))
192 			return -EPERM;
193 	} else if (!capable(CAP_SYS_RAWIO))
194 		return -EPERM;
195 
196 	/*
197 	 * fill in request structure
198 	 */
199 	rq->cmd_len = hdr->request_len;
200 	rq->cmd_type = REQ_TYPE_BLOCK_PC;
201 
202 	rq->timeout = (hdr->timeout * HZ) / 1000;
203 	if (!rq->timeout)
204 		rq->timeout = q->sg_timeout;
205 	if (!rq->timeout)
206 		rq->timeout = BLK_DEFAULT_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 	if (hdr->dout_xfer_len > (q->max_sectors << 9) ||
222 	    hdr->din_xfer_len > (q->max_sectors << 9))
223 		return -EIO;
224 
225 	switch (hdr->protocol) {
226 	case BSG_PROTOCOL_SCSI:
227 		switch (hdr->subprotocol) {
228 		case BSG_SUB_PROTOCOL_SCSI_CMD:
229 		case BSG_SUB_PROTOCOL_SCSI_TRANSPORT:
230 			break;
231 		default:
232 			ret = -EINVAL;
233 		}
234 		break;
235 	default:
236 		ret = -EINVAL;
237 	}
238 
239 	*rw = hdr->dout_xfer_len ? WRITE : READ;
240 	return ret;
241 }
242 
243 /*
244  * map sg_io_v4 to a request.
245  */
246 static struct request *
247 bsg_map_hdr(struct bsg_device *bd, struct sg_io_v4 *hdr)
248 {
249 	struct request_queue *q = bd->queue;
250 	struct request *rq, *next_rq = NULL;
251 	int ret, rw;
252 	unsigned int dxfer_len;
253 	void *dxferp = NULL;
254 
255 	dprintk("map hdr %llx/%u %llx/%u\n", (unsigned long long) hdr->dout_xferp,
256 		hdr->dout_xfer_len, (unsigned long long) hdr->din_xferp,
257 		hdr->din_xfer_len);
258 
259 	ret = bsg_validate_sgv4_hdr(q, hdr, &rw);
260 	if (ret)
261 		return ERR_PTR(ret);
262 
263 	/*
264 	 * map scatter-gather elements seperately and string them to request
265 	 */
266 	rq = blk_get_request(q, rw, GFP_KERNEL);
267 	if (!rq)
268 		return ERR_PTR(-ENOMEM);
269 	ret = blk_fill_sgv4_hdr_rq(q, rq, hdr, bd);
270 	if (ret)
271 		goto out;
272 
273 	if (rw == WRITE && hdr->din_xfer_len) {
274 		if (!test_bit(QUEUE_FLAG_BIDI, &q->queue_flags)) {
275 			ret = -EOPNOTSUPP;
276 			goto out;
277 		}
278 
279 		next_rq = blk_get_request(q, READ, GFP_KERNEL);
280 		if (!next_rq) {
281 			ret = -ENOMEM;
282 			goto out;
283 		}
284 		rq->next_rq = next_rq;
285 		next_rq->cmd_type = rq->cmd_type;
286 
287 		dxferp = (void*)(unsigned long)hdr->din_xferp;
288 		ret =  blk_rq_map_user(q, next_rq, dxferp, hdr->din_xfer_len);
289 		if (ret)
290 			goto out;
291 	}
292 
293 	if (hdr->dout_xfer_len) {
294 		dxfer_len = hdr->dout_xfer_len;
295 		dxferp = (void*)(unsigned long)hdr->dout_xferp;
296 	} else if (hdr->din_xfer_len) {
297 		dxfer_len = hdr->din_xfer_len;
298 		dxferp = (void*)(unsigned long)hdr->din_xferp;
299 	} else
300 		dxfer_len = 0;
301 
302 	if (dxfer_len) {
303 		ret = blk_rq_map_user(q, rq, dxferp, dxfer_len);
304 		if (ret)
305 			goto out;
306 	}
307 	return rq;
308 out:
309 	if (rq->cmd != rq->__cmd)
310 		kfree(rq->cmd);
311 	blk_put_request(rq);
312 	if (next_rq) {
313 		blk_rq_unmap_user(next_rq->bio);
314 		blk_put_request(next_rq);
315 	}
316 	return ERR_PTR(ret);
317 }
318 
319 /*
320  * async completion call-back from the block layer, when scsi/ide/whatever
321  * calls end_that_request_last() on a request
322  */
323 static void bsg_rq_end_io(struct request *rq, int uptodate)
324 {
325 	struct bsg_command *bc = rq->end_io_data;
326 	struct bsg_device *bd = bc->bd;
327 	unsigned long flags;
328 
329 	dprintk("%s: finished rq %p bc %p, bio %p stat %d\n",
330 		bd->name, rq, bc, bc->bio, uptodate);
331 
332 	bc->hdr.duration = jiffies_to_msecs(jiffies - bc->hdr.duration);
333 
334 	spin_lock_irqsave(&bd->lock, flags);
335 	list_move_tail(&bc->list, &bd->done_list);
336 	bd->done_cmds++;
337 	spin_unlock_irqrestore(&bd->lock, flags);
338 
339 	wake_up(&bd->wq_done);
340 }
341 
342 /*
343  * do final setup of a 'bc' and submit the matching 'rq' to the block
344  * layer for io
345  */
346 static void bsg_add_command(struct bsg_device *bd, struct request_queue *q,
347 			    struct bsg_command *bc, struct request *rq)
348 {
349 	rq->sense = bc->sense;
350 	rq->sense_len = 0;
351 
352 	/*
353 	 * add bc command to busy queue and submit rq for io
354 	 */
355 	bc->rq = rq;
356 	bc->bio = rq->bio;
357 	if (rq->next_rq)
358 		bc->bidi_bio = rq->next_rq->bio;
359 	bc->hdr.duration = jiffies;
360 	spin_lock_irq(&bd->lock);
361 	list_add_tail(&bc->list, &bd->busy_list);
362 	spin_unlock_irq(&bd->lock);
363 
364 	dprintk("%s: queueing rq %p, bc %p\n", bd->name, rq, bc);
365 
366 	rq->end_io_data = bc;
367 	blk_execute_rq_nowait(q, NULL, rq, 1, bsg_rq_end_io);
368 }
369 
370 static struct bsg_command *bsg_next_done_cmd(struct bsg_device *bd)
371 {
372 	struct bsg_command *bc = NULL;
373 
374 	spin_lock_irq(&bd->lock);
375 	if (bd->done_cmds) {
376 		bc = list_first_entry(&bd->done_list, struct bsg_command, list);
377 		list_del(&bc->list);
378 		bd->done_cmds--;
379 	}
380 	spin_unlock_irq(&bd->lock);
381 
382 	return bc;
383 }
384 
385 /*
386  * Get a finished command from the done list
387  */
388 static struct bsg_command *bsg_get_done_cmd(struct bsg_device *bd)
389 {
390 	struct bsg_command *bc;
391 	int ret;
392 
393 	do {
394 		bc = bsg_next_done_cmd(bd);
395 		if (bc)
396 			break;
397 
398 		if (!test_bit(BSG_F_BLOCK, &bd->flags)) {
399 			bc = ERR_PTR(-EAGAIN);
400 			break;
401 		}
402 
403 		ret = wait_event_interruptible(bd->wq_done, bd->done_cmds);
404 		if (ret) {
405 			bc = ERR_PTR(-ERESTARTSYS);
406 			break;
407 		}
408 	} while (1);
409 
410 	dprintk("%s: returning done %p\n", bd->name, bc);
411 
412 	return bc;
413 }
414 
415 static int blk_complete_sgv4_hdr_rq(struct request *rq, struct sg_io_v4 *hdr,
416 				    struct bio *bio, struct bio *bidi_bio)
417 {
418 	int ret = 0;
419 
420 	dprintk("rq %p bio %p %u\n", rq, bio, rq->errors);
421 	/*
422 	 * fill in all the output members
423 	 */
424 	hdr->device_status = status_byte(rq->errors);
425 	hdr->transport_status = host_byte(rq->errors);
426 	hdr->driver_status = driver_byte(rq->errors);
427 	hdr->info = 0;
428 	if (hdr->device_status || hdr->transport_status || hdr->driver_status)
429 		hdr->info |= SG_INFO_CHECK;
430 	hdr->response_len = 0;
431 
432 	if (rq->sense_len && hdr->response) {
433 		int len = min_t(unsigned int, hdr->max_response_len,
434 					rq->sense_len);
435 
436 		ret = copy_to_user((void*)(unsigned long)hdr->response,
437 				   rq->sense, len);
438 		if (!ret)
439 			hdr->response_len = len;
440 		else
441 			ret = -EFAULT;
442 	}
443 
444 	if (rq->next_rq) {
445 		hdr->dout_resid = rq->data_len;
446 		hdr->din_resid = rq->next_rq->data_len;
447 		blk_rq_unmap_user(bidi_bio);
448 		blk_put_request(rq->next_rq);
449 	} else if (rq_data_dir(rq) == READ)
450 		hdr->din_resid = rq->data_len;
451 	else
452 		hdr->dout_resid = rq->data_len;
453 
454 	/*
455 	 * If the request generated a negative error number, return it
456 	 * (providing we aren't already returning an error); if it's
457 	 * just a protocol response (i.e. non negative), that gets
458 	 * processed above.
459 	 */
460 	if (!ret && rq->errors < 0)
461 		ret = rq->errors;
462 
463 	blk_rq_unmap_user(bio);
464 	if (rq->cmd != rq->__cmd)
465 		kfree(rq->cmd);
466 	blk_put_request(rq);
467 
468 	return ret;
469 }
470 
471 static int bsg_complete_all_commands(struct bsg_device *bd)
472 {
473 	struct bsg_command *bc;
474 	int ret, tret;
475 
476 	dprintk("%s: entered\n", bd->name);
477 
478 	/*
479 	 * wait for all commands to complete
480 	 */
481 	ret = 0;
482 	do {
483 		ret = bsg_io_schedule(bd);
484 		/*
485 		 * look for -ENODATA specifically -- we'll sometimes get
486 		 * -ERESTARTSYS when we've taken a signal, but we can't
487 		 * return until we're done freeing the queue, so ignore
488 		 * it.  The signal will get handled when we're done freeing
489 		 * the bsg_device.
490 		 */
491 	} while (ret != -ENODATA);
492 
493 	/*
494 	 * discard done commands
495 	 */
496 	ret = 0;
497 	do {
498 		spin_lock_irq(&bd->lock);
499 		if (!bd->queued_cmds) {
500 			spin_unlock_irq(&bd->lock);
501 			break;
502 		}
503 		spin_unlock_irq(&bd->lock);
504 
505 		bc = bsg_get_done_cmd(bd);
506 		if (IS_ERR(bc))
507 			break;
508 
509 		tret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
510 						bc->bidi_bio);
511 		if (!ret)
512 			ret = tret;
513 
514 		bsg_free_command(bc);
515 	} while (1);
516 
517 	return ret;
518 }
519 
520 static int
521 __bsg_read(char __user *buf, size_t count, struct bsg_device *bd,
522 	   const struct iovec *iov, ssize_t *bytes_read)
523 {
524 	struct bsg_command *bc;
525 	int nr_commands, ret;
526 
527 	if (count % sizeof(struct sg_io_v4))
528 		return -EINVAL;
529 
530 	ret = 0;
531 	nr_commands = count / sizeof(struct sg_io_v4);
532 	while (nr_commands) {
533 		bc = bsg_get_done_cmd(bd);
534 		if (IS_ERR(bc)) {
535 			ret = PTR_ERR(bc);
536 			break;
537 		}
538 
539 		/*
540 		 * this is the only case where we need to copy data back
541 		 * after completing the request. so do that here,
542 		 * bsg_complete_work() cannot do that for us
543 		 */
544 		ret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
545 					       bc->bidi_bio);
546 
547 		if (copy_to_user(buf, &bc->hdr, sizeof(bc->hdr)))
548 			ret = -EFAULT;
549 
550 		bsg_free_command(bc);
551 
552 		if (ret)
553 			break;
554 
555 		buf += sizeof(struct sg_io_v4);
556 		*bytes_read += sizeof(struct sg_io_v4);
557 		nr_commands--;
558 	}
559 
560 	return ret;
561 }
562 
563 static inline void bsg_set_block(struct bsg_device *bd, struct file *file)
564 {
565 	if (file->f_flags & O_NONBLOCK)
566 		clear_bit(BSG_F_BLOCK, &bd->flags);
567 	else
568 		set_bit(BSG_F_BLOCK, &bd->flags);
569 }
570 
571 static void bsg_set_cmd_filter(struct bsg_device *bd,
572 			   struct file *file)
573 {
574 	struct inode *inode;
575 	struct gendisk *disk;
576 
577 	if (!file)
578 		return;
579 
580 	inode = file->f_dentry->d_inode;
581 	if (!inode)
582 		return;
583 
584 	disk = inode->i_bdev->bd_disk;
585 
586 	bd->cmd_filter = &disk->cmd_filter;
587 	bd->f_mode = &file->f_mode;
588 }
589 
590 /*
591  * Check if the error is a "real" error that we should return.
592  */
593 static inline int err_block_err(int ret)
594 {
595 	if (ret && ret != -ENOSPC && ret != -ENODATA && ret != -EAGAIN)
596 		return 1;
597 
598 	return 0;
599 }
600 
601 static ssize_t
602 bsg_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
603 {
604 	struct bsg_device *bd = file->private_data;
605 	int ret;
606 	ssize_t bytes_read;
607 
608 	dprintk("%s: read %Zd bytes\n", bd->name, count);
609 
610 	bsg_set_block(bd, file);
611 	bsg_set_cmd_filter(bd, file);
612 
613 	bytes_read = 0;
614 	ret = __bsg_read(buf, count, bd, NULL, &bytes_read);
615 	*ppos = bytes_read;
616 
617 	if (!bytes_read || (bytes_read && err_block_err(ret)))
618 		bytes_read = ret;
619 
620 	return bytes_read;
621 }
622 
623 static int __bsg_write(struct bsg_device *bd, const char __user *buf,
624 		       size_t count, ssize_t *bytes_written)
625 {
626 	struct bsg_command *bc;
627 	struct request *rq;
628 	int ret, nr_commands;
629 
630 	if (count % sizeof(struct sg_io_v4))
631 		return -EINVAL;
632 
633 	nr_commands = count / sizeof(struct sg_io_v4);
634 	rq = NULL;
635 	bc = NULL;
636 	ret = 0;
637 	while (nr_commands) {
638 		struct request_queue *q = bd->queue;
639 
640 		bc = bsg_alloc_command(bd);
641 		if (IS_ERR(bc)) {
642 			ret = PTR_ERR(bc);
643 			bc = NULL;
644 			break;
645 		}
646 
647 		if (copy_from_user(&bc->hdr, buf, sizeof(bc->hdr))) {
648 			ret = -EFAULT;
649 			break;
650 		}
651 
652 		/*
653 		 * get a request, fill in the blanks, and add to request queue
654 		 */
655 		rq = bsg_map_hdr(bd, &bc->hdr);
656 		if (IS_ERR(rq)) {
657 			ret = PTR_ERR(rq);
658 			rq = NULL;
659 			break;
660 		}
661 
662 		bsg_add_command(bd, q, bc, rq);
663 		bc = NULL;
664 		rq = NULL;
665 		nr_commands--;
666 		buf += sizeof(struct sg_io_v4);
667 		*bytes_written += sizeof(struct sg_io_v4);
668 	}
669 
670 	if (bc)
671 		bsg_free_command(bc);
672 
673 	return ret;
674 }
675 
676 static ssize_t
677 bsg_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
678 {
679 	struct bsg_device *bd = file->private_data;
680 	ssize_t bytes_written;
681 	int ret;
682 
683 	dprintk("%s: write %Zd bytes\n", bd->name, count);
684 
685 	bsg_set_block(bd, file);
686 	bsg_set_cmd_filter(bd, file);
687 
688 	bytes_written = 0;
689 	ret = __bsg_write(bd, buf, count, &bytes_written);
690 	*ppos = bytes_written;
691 
692 	/*
693 	 * return bytes written on non-fatal errors
694 	 */
695 	if (!bytes_written || (bytes_written && err_block_err(ret)))
696 		bytes_written = ret;
697 
698 	dprintk("%s: returning %Zd\n", bd->name, bytes_written);
699 	return bytes_written;
700 }
701 
702 static struct bsg_device *bsg_alloc_device(void)
703 {
704 	struct bsg_device *bd;
705 
706 	bd = kzalloc(sizeof(struct bsg_device), GFP_KERNEL);
707 	if (unlikely(!bd))
708 		return NULL;
709 
710 	spin_lock_init(&bd->lock);
711 
712 	bd->max_queue = BSG_DEFAULT_CMDS;
713 
714 	INIT_LIST_HEAD(&bd->busy_list);
715 	INIT_LIST_HEAD(&bd->done_list);
716 	INIT_HLIST_NODE(&bd->dev_list);
717 
718 	init_waitqueue_head(&bd->wq_free);
719 	init_waitqueue_head(&bd->wq_done);
720 	return bd;
721 }
722 
723 static void bsg_kref_release_function(struct kref *kref)
724 {
725 	struct bsg_class_device *bcd =
726 		container_of(kref, struct bsg_class_device, ref);
727 	struct device *parent = bcd->parent;
728 
729 	if (bcd->release)
730 		bcd->release(bcd->parent);
731 
732 	put_device(parent);
733 }
734 
735 static int bsg_put_device(struct bsg_device *bd)
736 {
737 	int ret = 0, do_free;
738 	struct request_queue *q = bd->queue;
739 
740 	mutex_lock(&bsg_mutex);
741 
742 	do_free = atomic_dec_and_test(&bd->ref_count);
743 	if (!do_free) {
744 		mutex_unlock(&bsg_mutex);
745 		goto out;
746 	}
747 
748 	hlist_del(&bd->dev_list);
749 	mutex_unlock(&bsg_mutex);
750 
751 	dprintk("%s: tearing down\n", bd->name);
752 
753 	/*
754 	 * close can always block
755 	 */
756 	set_bit(BSG_F_BLOCK, &bd->flags);
757 
758 	/*
759 	 * correct error detection baddies here again. it's the responsibility
760 	 * of the app to properly reap commands before close() if it wants
761 	 * fool-proof error detection
762 	 */
763 	ret = bsg_complete_all_commands(bd);
764 
765 	kfree(bd);
766 out:
767 	kref_put(&q->bsg_dev.ref, bsg_kref_release_function);
768 	if (do_free)
769 		blk_put_queue(q);
770 	return ret;
771 }
772 
773 static struct bsg_device *bsg_add_device(struct inode *inode,
774 					 struct request_queue *rq,
775 					 struct file *file)
776 {
777 	struct bsg_device *bd;
778 	int ret;
779 #ifdef BSG_DEBUG
780 	unsigned char buf[32];
781 #endif
782 	ret = blk_get_queue(rq);
783 	if (ret)
784 		return ERR_PTR(-ENXIO);
785 
786 	bd = bsg_alloc_device();
787 	if (!bd) {
788 		blk_put_queue(rq);
789 		return ERR_PTR(-ENOMEM);
790 	}
791 
792 	bd->queue = rq;
793 
794 	bsg_set_block(bd, file);
795 	bsg_set_cmd_filter(bd, file);
796 
797 	atomic_set(&bd->ref_count, 1);
798 	mutex_lock(&bsg_mutex);
799 	hlist_add_head(&bd->dev_list, bsg_dev_idx_hash(iminor(inode)));
800 
801 	strncpy(bd->name, rq->bsg_dev.class_dev->bus_id, sizeof(bd->name) - 1);
802 	dprintk("bound to <%s>, max queue %d\n",
803 		format_dev_t(buf, inode->i_rdev), bd->max_queue);
804 
805 	mutex_unlock(&bsg_mutex);
806 	return bd;
807 }
808 
809 static struct bsg_device *__bsg_get_device(int minor, struct request_queue *q)
810 {
811 	struct bsg_device *bd;
812 	struct hlist_node *entry;
813 
814 	mutex_lock(&bsg_mutex);
815 
816 	hlist_for_each_entry(bd, entry, bsg_dev_idx_hash(minor), dev_list) {
817 		if (bd->queue == q) {
818 			atomic_inc(&bd->ref_count);
819 			goto found;
820 		}
821 	}
822 	bd = NULL;
823 found:
824 	mutex_unlock(&bsg_mutex);
825 	return bd;
826 }
827 
828 static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file)
829 {
830 	struct bsg_device *bd;
831 	struct bsg_class_device *bcd;
832 
833 	/*
834 	 * find the class device
835 	 */
836 	mutex_lock(&bsg_mutex);
837 	bcd = idr_find(&bsg_minor_idr, iminor(inode));
838 	if (bcd)
839 		kref_get(&bcd->ref);
840 	mutex_unlock(&bsg_mutex);
841 
842 	if (!bcd)
843 		return ERR_PTR(-ENODEV);
844 
845 	bd = __bsg_get_device(iminor(inode), bcd->queue);
846 	if (bd)
847 		return bd;
848 
849 	bd = bsg_add_device(inode, bcd->queue, file);
850 	if (IS_ERR(bd))
851 		kref_put(&bcd->ref, bsg_kref_release_function);
852 
853 	return bd;
854 }
855 
856 static int bsg_open(struct inode *inode, struct file *file)
857 {
858 	struct bsg_device *bd;
859 
860 	lock_kernel();
861 	bd = bsg_get_device(inode, file);
862 	unlock_kernel();
863 
864 	if (IS_ERR(bd))
865 		return PTR_ERR(bd);
866 
867 	file->private_data = bd;
868 	return 0;
869 }
870 
871 static int bsg_release(struct inode *inode, struct file *file)
872 {
873 	struct bsg_device *bd = file->private_data;
874 
875 	file->private_data = NULL;
876 	return bsg_put_device(bd);
877 }
878 
879 static unsigned int bsg_poll(struct file *file, poll_table *wait)
880 {
881 	struct bsg_device *bd = file->private_data;
882 	unsigned int mask = 0;
883 
884 	poll_wait(file, &bd->wq_done, wait);
885 	poll_wait(file, &bd->wq_free, wait);
886 
887 	spin_lock_irq(&bd->lock);
888 	if (!list_empty(&bd->done_list))
889 		mask |= POLLIN | POLLRDNORM;
890 	if (bd->queued_cmds >= bd->max_queue)
891 		mask |= POLLOUT;
892 	spin_unlock_irq(&bd->lock);
893 
894 	return mask;
895 }
896 
897 static long bsg_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
898 {
899 	struct bsg_device *bd = file->private_data;
900 	int __user *uarg = (int __user *) arg;
901 	int ret;
902 
903 	switch (cmd) {
904 		/*
905 		 * our own ioctls
906 		 */
907 	case SG_GET_COMMAND_Q:
908 		return put_user(bd->max_queue, uarg);
909 	case SG_SET_COMMAND_Q: {
910 		int queue;
911 
912 		if (get_user(queue, uarg))
913 			return -EFAULT;
914 		if (queue < 1)
915 			return -EINVAL;
916 
917 		spin_lock_irq(&bd->lock);
918 		bd->max_queue = queue;
919 		spin_unlock_irq(&bd->lock);
920 		return 0;
921 	}
922 
923 	/*
924 	 * SCSI/sg ioctls
925 	 */
926 	case SG_GET_VERSION_NUM:
927 	case SCSI_IOCTL_GET_IDLUN:
928 	case SCSI_IOCTL_GET_BUS_NUMBER:
929 	case SG_SET_TIMEOUT:
930 	case SG_GET_TIMEOUT:
931 	case SG_GET_RESERVED_SIZE:
932 	case SG_SET_RESERVED_SIZE:
933 	case SG_EMULATED_HOST:
934 	case SCSI_IOCTL_SEND_COMMAND: {
935 		void __user *uarg = (void __user *) arg;
936 		return scsi_cmd_ioctl(file, bd->queue, NULL, cmd, uarg);
937 	}
938 	case SG_IO: {
939 		struct request *rq;
940 		struct bio *bio, *bidi_bio = NULL;
941 		struct sg_io_v4 hdr;
942 
943 		if (copy_from_user(&hdr, uarg, sizeof(hdr)))
944 			return -EFAULT;
945 
946 		rq = bsg_map_hdr(bd, &hdr);
947 		if (IS_ERR(rq))
948 			return PTR_ERR(rq);
949 
950 		bio = rq->bio;
951 		if (rq->next_rq)
952 			bidi_bio = rq->next_rq->bio;
953 		blk_execute_rq(bd->queue, NULL, rq, 0);
954 		ret = blk_complete_sgv4_hdr_rq(rq, &hdr, bio, bidi_bio);
955 
956 		if (copy_to_user(uarg, &hdr, sizeof(hdr)))
957 			return -EFAULT;
958 
959 		return ret;
960 	}
961 	/*
962 	 * block device ioctls
963 	 */
964 	default:
965 #if 0
966 		return ioctl_by_bdev(bd->bdev, cmd, arg);
967 #else
968 		return -ENOTTY;
969 #endif
970 	}
971 }
972 
973 static const struct file_operations bsg_fops = {
974 	.read		=	bsg_read,
975 	.write		=	bsg_write,
976 	.poll		=	bsg_poll,
977 	.open		=	bsg_open,
978 	.release	=	bsg_release,
979 	.unlocked_ioctl	=	bsg_ioctl,
980 	.owner		=	THIS_MODULE,
981 };
982 
983 void bsg_unregister_queue(struct request_queue *q)
984 {
985 	struct bsg_class_device *bcd = &q->bsg_dev;
986 
987 	if (!bcd->class_dev)
988 		return;
989 
990 	mutex_lock(&bsg_mutex);
991 	idr_remove(&bsg_minor_idr, bcd->minor);
992 	sysfs_remove_link(&q->kobj, "bsg");
993 	device_unregister(bcd->class_dev);
994 	bcd->class_dev = NULL;
995 	kref_put(&bcd->ref, bsg_kref_release_function);
996 	mutex_unlock(&bsg_mutex);
997 }
998 EXPORT_SYMBOL_GPL(bsg_unregister_queue);
999 
1000 int bsg_register_queue(struct request_queue *q, struct device *parent,
1001 		       const char *name, void (*release)(struct device *))
1002 {
1003 	struct bsg_class_device *bcd;
1004 	dev_t dev;
1005 	int ret, minor;
1006 	struct device *class_dev = NULL;
1007 	const char *devname;
1008 
1009 	if (name)
1010 		devname = name;
1011 	else
1012 		devname = parent->bus_id;
1013 
1014 	/*
1015 	 * we need a proper transport to send commands, not a stacked device
1016 	 */
1017 	if (!q->request_fn)
1018 		return 0;
1019 
1020 	bcd = &q->bsg_dev;
1021 	memset(bcd, 0, sizeof(*bcd));
1022 
1023 	mutex_lock(&bsg_mutex);
1024 
1025 	ret = idr_pre_get(&bsg_minor_idr, GFP_KERNEL);
1026 	if (!ret) {
1027 		ret = -ENOMEM;
1028 		goto unlock;
1029 	}
1030 
1031 	ret = idr_get_new(&bsg_minor_idr, bcd, &minor);
1032 	if (ret < 0)
1033 		goto unlock;
1034 
1035 	if (minor >= BSG_MAX_DEVS) {
1036 		printk(KERN_ERR "bsg: too many bsg devices\n");
1037 		ret = -EINVAL;
1038 		goto remove_idr;
1039 	}
1040 
1041 	bcd->minor = minor;
1042 	bcd->queue = q;
1043 	bcd->parent = get_device(parent);
1044 	bcd->release = release;
1045 	kref_init(&bcd->ref);
1046 	dev = MKDEV(bsg_major, bcd->minor);
1047 	class_dev = device_create(bsg_class, parent, dev, "%s", devname);
1048 	if (IS_ERR(class_dev)) {
1049 		ret = PTR_ERR(class_dev);
1050 		goto put_dev;
1051 	}
1052 	bcd->class_dev = class_dev;
1053 
1054 	if (q->kobj.sd) {
1055 		ret = sysfs_create_link(&q->kobj, &bcd->class_dev->kobj, "bsg");
1056 		if (ret)
1057 			goto unregister_class_dev;
1058 	}
1059 
1060 	mutex_unlock(&bsg_mutex);
1061 	return 0;
1062 
1063 unregister_class_dev:
1064 	device_unregister(class_dev);
1065 put_dev:
1066 	put_device(parent);
1067 remove_idr:
1068 	idr_remove(&bsg_minor_idr, minor);
1069 unlock:
1070 	mutex_unlock(&bsg_mutex);
1071 	return ret;
1072 }
1073 EXPORT_SYMBOL_GPL(bsg_register_queue);
1074 
1075 static struct cdev bsg_cdev;
1076 
1077 static int __init bsg_init(void)
1078 {
1079 	int ret, i;
1080 	dev_t devid;
1081 
1082 	bsg_cmd_cachep = kmem_cache_create("bsg_cmd",
1083 				sizeof(struct bsg_command), 0, 0, NULL);
1084 	if (!bsg_cmd_cachep) {
1085 		printk(KERN_ERR "bsg: failed creating slab cache\n");
1086 		return -ENOMEM;
1087 	}
1088 
1089 	for (i = 0; i < BSG_LIST_ARRAY_SIZE; i++)
1090 		INIT_HLIST_HEAD(&bsg_device_list[i]);
1091 
1092 	bsg_class = class_create(THIS_MODULE, "bsg");
1093 	if (IS_ERR(bsg_class)) {
1094 		ret = PTR_ERR(bsg_class);
1095 		goto destroy_kmemcache;
1096 	}
1097 
1098 	ret = alloc_chrdev_region(&devid, 0, BSG_MAX_DEVS, "bsg");
1099 	if (ret)
1100 		goto destroy_bsg_class;
1101 
1102 	bsg_major = MAJOR(devid);
1103 
1104 	cdev_init(&bsg_cdev, &bsg_fops);
1105 	ret = cdev_add(&bsg_cdev, MKDEV(bsg_major, 0), BSG_MAX_DEVS);
1106 	if (ret)
1107 		goto unregister_chrdev;
1108 
1109 	printk(KERN_INFO BSG_DESCRIPTION " version " BSG_VERSION
1110 	       " loaded (major %d)\n", bsg_major);
1111 	return 0;
1112 unregister_chrdev:
1113 	unregister_chrdev_region(MKDEV(bsg_major, 0), BSG_MAX_DEVS);
1114 destroy_bsg_class:
1115 	class_destroy(bsg_class);
1116 destroy_kmemcache:
1117 	kmem_cache_destroy(bsg_cmd_cachep);
1118 	return ret;
1119 }
1120 
1121 MODULE_AUTHOR("Jens Axboe");
1122 MODULE_DESCRIPTION(BSG_DESCRIPTION);
1123 MODULE_LICENSE("GPL");
1124 
1125 device_initcall(bsg_init);
1126