xref: /linux/block/bsg.c (revision 156f5a7801195fa2ce44aeeb62d6cf8468f3332a)
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(&q->cmd_filter, 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 		next_rq->bio = NULL;
319 		blk_put_request(next_rq);
320 	}
321 	return ERR_PTR(ret);
322 }
323 
324 /*
325  * async completion call-back from the block layer, when scsi/ide/whatever
326  * calls end_that_request_last() on a request
327  */
328 static void bsg_rq_end_io(struct request *rq, int uptodate)
329 {
330 	struct bsg_command *bc = rq->end_io_data;
331 	struct bsg_device *bd = bc->bd;
332 	unsigned long flags;
333 
334 	dprintk("%s: finished rq %p bc %p, bio %p stat %d\n",
335 		bd->name, rq, bc, bc->bio, uptodate);
336 
337 	bc->hdr.duration = jiffies_to_msecs(jiffies - bc->hdr.duration);
338 
339 	spin_lock_irqsave(&bd->lock, flags);
340 	list_move_tail(&bc->list, &bd->done_list);
341 	bd->done_cmds++;
342 	spin_unlock_irqrestore(&bd->lock, flags);
343 
344 	wake_up(&bd->wq_done);
345 }
346 
347 /*
348  * do final setup of a 'bc' and submit the matching 'rq' to the block
349  * layer for io
350  */
351 static void bsg_add_command(struct bsg_device *bd, struct request_queue *q,
352 			    struct bsg_command *bc, struct request *rq)
353 {
354 	int at_head = (0 == (bc->hdr.flags & BSG_FLAG_Q_AT_TAIL));
355 
356 	/*
357 	 * add bc command to busy queue and submit rq for io
358 	 */
359 	bc->rq = rq;
360 	bc->bio = rq->bio;
361 	if (rq->next_rq)
362 		bc->bidi_bio = rq->next_rq->bio;
363 	bc->hdr.duration = jiffies;
364 	spin_lock_irq(&bd->lock);
365 	list_add_tail(&bc->list, &bd->busy_list);
366 	spin_unlock_irq(&bd->lock);
367 
368 	dprintk("%s: queueing rq %p, bc %p\n", bd->name, rq, bc);
369 
370 	rq->end_io_data = bc;
371 	blk_execute_rq_nowait(q, NULL, rq, at_head, bsg_rq_end_io);
372 }
373 
374 static struct bsg_command *bsg_next_done_cmd(struct bsg_device *bd)
375 {
376 	struct bsg_command *bc = NULL;
377 
378 	spin_lock_irq(&bd->lock);
379 	if (bd->done_cmds) {
380 		bc = list_first_entry(&bd->done_list, struct bsg_command, list);
381 		list_del(&bc->list);
382 		bd->done_cmds--;
383 	}
384 	spin_unlock_irq(&bd->lock);
385 
386 	return bc;
387 }
388 
389 /*
390  * Get a finished command from the done list
391  */
392 static struct bsg_command *bsg_get_done_cmd(struct bsg_device *bd)
393 {
394 	struct bsg_command *bc;
395 	int ret;
396 
397 	do {
398 		bc = bsg_next_done_cmd(bd);
399 		if (bc)
400 			break;
401 
402 		if (!test_bit(BSG_F_BLOCK, &bd->flags)) {
403 			bc = ERR_PTR(-EAGAIN);
404 			break;
405 		}
406 
407 		ret = wait_event_interruptible(bd->wq_done, bd->done_cmds);
408 		if (ret) {
409 			bc = ERR_PTR(-ERESTARTSYS);
410 			break;
411 		}
412 	} while (1);
413 
414 	dprintk("%s: returning done %p\n", bd->name, bc);
415 
416 	return bc;
417 }
418 
419 static int blk_complete_sgv4_hdr_rq(struct request *rq, struct sg_io_v4 *hdr,
420 				    struct bio *bio, struct bio *bidi_bio)
421 {
422 	int ret = 0;
423 
424 	dprintk("rq %p bio %p 0x%x\n", rq, bio, rq->errors);
425 	/*
426 	 * fill in all the output members
427 	 */
428 	hdr->device_status = status_byte(rq->errors);
429 	hdr->transport_status = host_byte(rq->errors);
430 	hdr->driver_status = driver_byte(rq->errors);
431 	hdr->info = 0;
432 	if (hdr->device_status || hdr->transport_status || hdr->driver_status)
433 		hdr->info |= SG_INFO_CHECK;
434 	hdr->response_len = 0;
435 
436 	if (rq->sense_len && hdr->response) {
437 		int len = min_t(unsigned int, hdr->max_response_len,
438 					rq->sense_len);
439 
440 		ret = copy_to_user((void*)(unsigned long)hdr->response,
441 				   rq->sense, len);
442 		if (!ret)
443 			hdr->response_len = len;
444 		else
445 			ret = -EFAULT;
446 	}
447 
448 	if (rq->next_rq) {
449 		hdr->dout_resid = rq->resid_len;
450 		hdr->din_resid = rq->next_rq->resid_len;
451 		blk_rq_unmap_user(bidi_bio);
452 		rq->next_rq->bio = NULL;
453 		blk_put_request(rq->next_rq);
454 	} else if (rq_data_dir(rq) == READ)
455 		hdr->din_resid = rq->resid_len;
456 	else
457 		hdr->dout_resid = rq->resid_len;
458 
459 	/*
460 	 * If the request generated a negative error number, return it
461 	 * (providing we aren't already returning an error); if it's
462 	 * just a protocol response (i.e. non negative), that gets
463 	 * processed above.
464 	 */
465 	if (!ret && rq->errors < 0)
466 		ret = rq->errors;
467 
468 	blk_rq_unmap_user(bio);
469 	if (rq->cmd != rq->__cmd)
470 		kfree(rq->cmd);
471 	rq->bio = NULL;
472 	blk_put_request(rq);
473 
474 	return ret;
475 }
476 
477 static int bsg_complete_all_commands(struct bsg_device *bd)
478 {
479 	struct bsg_command *bc;
480 	int ret, tret;
481 
482 	dprintk("%s: entered\n", bd->name);
483 
484 	/*
485 	 * wait for all commands to complete
486 	 */
487 	ret = 0;
488 	do {
489 		ret = bsg_io_schedule(bd);
490 		/*
491 		 * look for -ENODATA specifically -- we'll sometimes get
492 		 * -ERESTARTSYS when we've taken a signal, but we can't
493 		 * return until we're done freeing the queue, so ignore
494 		 * it.  The signal will get handled when we're done freeing
495 		 * the bsg_device.
496 		 */
497 	} while (ret != -ENODATA);
498 
499 	/*
500 	 * discard done commands
501 	 */
502 	ret = 0;
503 	do {
504 		spin_lock_irq(&bd->lock);
505 		if (!bd->queued_cmds) {
506 			spin_unlock_irq(&bd->lock);
507 			break;
508 		}
509 		spin_unlock_irq(&bd->lock);
510 
511 		bc = bsg_get_done_cmd(bd);
512 		if (IS_ERR(bc))
513 			break;
514 
515 		tret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
516 						bc->bidi_bio);
517 		if (!ret)
518 			ret = tret;
519 
520 		bsg_free_command(bc);
521 	} while (1);
522 
523 	return ret;
524 }
525 
526 static int
527 __bsg_read(char __user *buf, size_t count, struct bsg_device *bd,
528 	   const struct iovec *iov, ssize_t *bytes_read)
529 {
530 	struct bsg_command *bc;
531 	int nr_commands, ret;
532 
533 	if (count % sizeof(struct sg_io_v4))
534 		return -EINVAL;
535 
536 	ret = 0;
537 	nr_commands = count / sizeof(struct sg_io_v4);
538 	while (nr_commands) {
539 		bc = bsg_get_done_cmd(bd);
540 		if (IS_ERR(bc)) {
541 			ret = PTR_ERR(bc);
542 			break;
543 		}
544 
545 		/*
546 		 * this is the only case where we need to copy data back
547 		 * after completing the request. so do that here,
548 		 * bsg_complete_work() cannot do that for us
549 		 */
550 		ret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
551 					       bc->bidi_bio);
552 
553 		if (copy_to_user(buf, &bc->hdr, sizeof(bc->hdr)))
554 			ret = -EFAULT;
555 
556 		bsg_free_command(bc);
557 
558 		if (ret)
559 			break;
560 
561 		buf += sizeof(struct sg_io_v4);
562 		*bytes_read += sizeof(struct sg_io_v4);
563 		nr_commands--;
564 	}
565 
566 	return ret;
567 }
568 
569 static inline void bsg_set_block(struct bsg_device *bd, struct file *file)
570 {
571 	if (file->f_flags & O_NONBLOCK)
572 		clear_bit(BSG_F_BLOCK, &bd->flags);
573 	else
574 		set_bit(BSG_F_BLOCK, &bd->flags);
575 }
576 
577 /*
578  * Check if the error is a "real" error that we should return.
579  */
580 static inline int err_block_err(int ret)
581 {
582 	if (ret && ret != -ENOSPC && ret != -ENODATA && ret != -EAGAIN)
583 		return 1;
584 
585 	return 0;
586 }
587 
588 static ssize_t
589 bsg_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
590 {
591 	struct bsg_device *bd = file->private_data;
592 	int ret;
593 	ssize_t bytes_read;
594 
595 	dprintk("%s: read %Zd bytes\n", bd->name, count);
596 
597 	bsg_set_block(bd, file);
598 
599 	bytes_read = 0;
600 	ret = __bsg_read(buf, count, bd, NULL, &bytes_read);
601 	*ppos = bytes_read;
602 
603 	if (!bytes_read || (bytes_read && err_block_err(ret)))
604 		bytes_read = ret;
605 
606 	return bytes_read;
607 }
608 
609 static int __bsg_write(struct bsg_device *bd, const char __user *buf,
610 		       size_t count, ssize_t *bytes_written,
611 		       fmode_t has_write_perm)
612 {
613 	struct bsg_command *bc;
614 	struct request *rq;
615 	int ret, nr_commands;
616 
617 	if (count % sizeof(struct sg_io_v4))
618 		return -EINVAL;
619 
620 	nr_commands = count / sizeof(struct sg_io_v4);
621 	rq = NULL;
622 	bc = NULL;
623 	ret = 0;
624 	while (nr_commands) {
625 		struct request_queue *q = bd->queue;
626 
627 		bc = bsg_alloc_command(bd);
628 		if (IS_ERR(bc)) {
629 			ret = PTR_ERR(bc);
630 			bc = NULL;
631 			break;
632 		}
633 
634 		if (copy_from_user(&bc->hdr, buf, sizeof(bc->hdr))) {
635 			ret = -EFAULT;
636 			break;
637 		}
638 
639 		/*
640 		 * get a request, fill in the blanks, and add to request queue
641 		 */
642 		rq = bsg_map_hdr(bd, &bc->hdr, has_write_perm, bc->sense);
643 		if (IS_ERR(rq)) {
644 			ret = PTR_ERR(rq);
645 			rq = NULL;
646 			break;
647 		}
648 
649 		bsg_add_command(bd, q, bc, rq);
650 		bc = NULL;
651 		rq = NULL;
652 		nr_commands--;
653 		buf += sizeof(struct sg_io_v4);
654 		*bytes_written += sizeof(struct sg_io_v4);
655 	}
656 
657 	if (bc)
658 		bsg_free_command(bc);
659 
660 	return ret;
661 }
662 
663 static ssize_t
664 bsg_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
665 {
666 	struct bsg_device *bd = file->private_data;
667 	ssize_t bytes_written;
668 	int ret;
669 
670 	dprintk("%s: write %Zd bytes\n", bd->name, count);
671 
672 	bsg_set_block(bd, file);
673 
674 	bytes_written = 0;
675 	ret = __bsg_write(bd, buf, count, &bytes_written,
676 			  file->f_mode & FMODE_WRITE);
677 
678 	*ppos = bytes_written;
679 
680 	/*
681 	 * return bytes written on non-fatal errors
682 	 */
683 	if (!bytes_written || (bytes_written && err_block_err(ret)))
684 		bytes_written = ret;
685 
686 	dprintk("%s: returning %Zd\n", bd->name, bytes_written);
687 	return bytes_written;
688 }
689 
690 static struct bsg_device *bsg_alloc_device(void)
691 {
692 	struct bsg_device *bd;
693 
694 	bd = kzalloc(sizeof(struct bsg_device), GFP_KERNEL);
695 	if (unlikely(!bd))
696 		return NULL;
697 
698 	spin_lock_init(&bd->lock);
699 
700 	bd->max_queue = BSG_DEFAULT_CMDS;
701 
702 	INIT_LIST_HEAD(&bd->busy_list);
703 	INIT_LIST_HEAD(&bd->done_list);
704 	INIT_HLIST_NODE(&bd->dev_list);
705 
706 	init_waitqueue_head(&bd->wq_free);
707 	init_waitqueue_head(&bd->wq_done);
708 	return bd;
709 }
710 
711 static void bsg_kref_release_function(struct kref *kref)
712 {
713 	struct bsg_class_device *bcd =
714 		container_of(kref, struct bsg_class_device, ref);
715 	struct device *parent = bcd->parent;
716 
717 	if (bcd->release)
718 		bcd->release(bcd->parent);
719 
720 	put_device(parent);
721 }
722 
723 static int bsg_put_device(struct bsg_device *bd)
724 {
725 	int ret = 0, do_free;
726 	struct request_queue *q = bd->queue;
727 
728 	mutex_lock(&bsg_mutex);
729 
730 	do_free = atomic_dec_and_test(&bd->ref_count);
731 	if (!do_free) {
732 		mutex_unlock(&bsg_mutex);
733 		goto out;
734 	}
735 
736 	hlist_del(&bd->dev_list);
737 	mutex_unlock(&bsg_mutex);
738 
739 	dprintk("%s: tearing down\n", bd->name);
740 
741 	/*
742 	 * close can always block
743 	 */
744 	set_bit(BSG_F_BLOCK, &bd->flags);
745 
746 	/*
747 	 * correct error detection baddies here again. it's the responsibility
748 	 * of the app to properly reap commands before close() if it wants
749 	 * fool-proof error detection
750 	 */
751 	ret = bsg_complete_all_commands(bd);
752 
753 	kfree(bd);
754 out:
755 	kref_put(&q->bsg_dev.ref, bsg_kref_release_function);
756 	if (do_free)
757 		blk_put_queue(q);
758 	return ret;
759 }
760 
761 static struct bsg_device *bsg_add_device(struct inode *inode,
762 					 struct request_queue *rq,
763 					 struct file *file)
764 {
765 	struct bsg_device *bd;
766 	int ret;
767 #ifdef BSG_DEBUG
768 	unsigned char buf[32];
769 #endif
770 	ret = blk_get_queue(rq);
771 	if (ret)
772 		return ERR_PTR(-ENXIO);
773 
774 	bd = bsg_alloc_device();
775 	if (!bd) {
776 		blk_put_queue(rq);
777 		return ERR_PTR(-ENOMEM);
778 	}
779 
780 	bd->queue = rq;
781 
782 	bsg_set_block(bd, file);
783 
784 	atomic_set(&bd->ref_count, 1);
785 	mutex_lock(&bsg_mutex);
786 	hlist_add_head(&bd->dev_list, bsg_dev_idx_hash(iminor(inode)));
787 
788 	strncpy(bd->name, dev_name(rq->bsg_dev.class_dev), sizeof(bd->name) - 1);
789 	dprintk("bound to <%s>, max queue %d\n",
790 		format_dev_t(buf, inode->i_rdev), bd->max_queue);
791 
792 	mutex_unlock(&bsg_mutex);
793 	return bd;
794 }
795 
796 static struct bsg_device *__bsg_get_device(int minor, struct request_queue *q)
797 {
798 	struct bsg_device *bd;
799 	struct hlist_node *entry;
800 
801 	mutex_lock(&bsg_mutex);
802 
803 	hlist_for_each_entry(bd, entry, bsg_dev_idx_hash(minor), dev_list) {
804 		if (bd->queue == q) {
805 			atomic_inc(&bd->ref_count);
806 			goto found;
807 		}
808 	}
809 	bd = NULL;
810 found:
811 	mutex_unlock(&bsg_mutex);
812 	return bd;
813 }
814 
815 static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file)
816 {
817 	struct bsg_device *bd;
818 	struct bsg_class_device *bcd;
819 
820 	/*
821 	 * find the class device
822 	 */
823 	mutex_lock(&bsg_mutex);
824 	bcd = idr_find(&bsg_minor_idr, iminor(inode));
825 	if (bcd)
826 		kref_get(&bcd->ref);
827 	mutex_unlock(&bsg_mutex);
828 
829 	if (!bcd)
830 		return ERR_PTR(-ENODEV);
831 
832 	bd = __bsg_get_device(iminor(inode), bcd->queue);
833 	if (bd)
834 		return bd;
835 
836 	bd = bsg_add_device(inode, bcd->queue, file);
837 	if (IS_ERR(bd))
838 		kref_put(&bcd->ref, bsg_kref_release_function);
839 
840 	return bd;
841 }
842 
843 static int bsg_open(struct inode *inode, struct file *file)
844 {
845 	struct bsg_device *bd;
846 
847 	lock_kernel();
848 	bd = bsg_get_device(inode, file);
849 	unlock_kernel();
850 
851 	if (IS_ERR(bd))
852 		return PTR_ERR(bd);
853 
854 	file->private_data = bd;
855 	return 0;
856 }
857 
858 static int bsg_release(struct inode *inode, struct file *file)
859 {
860 	struct bsg_device *bd = file->private_data;
861 
862 	file->private_data = NULL;
863 	return bsg_put_device(bd);
864 }
865 
866 static unsigned int bsg_poll(struct file *file, poll_table *wait)
867 {
868 	struct bsg_device *bd = file->private_data;
869 	unsigned int mask = 0;
870 
871 	poll_wait(file, &bd->wq_done, wait);
872 	poll_wait(file, &bd->wq_free, wait);
873 
874 	spin_lock_irq(&bd->lock);
875 	if (!list_empty(&bd->done_list))
876 		mask |= POLLIN | POLLRDNORM;
877 	if (bd->queued_cmds >= bd->max_queue)
878 		mask |= POLLOUT;
879 	spin_unlock_irq(&bd->lock);
880 
881 	return mask;
882 }
883 
884 static long bsg_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
885 {
886 	struct bsg_device *bd = file->private_data;
887 	int __user *uarg = (int __user *) arg;
888 	int ret;
889 
890 	switch (cmd) {
891 		/*
892 		 * our own ioctls
893 		 */
894 	case SG_GET_COMMAND_Q:
895 		return put_user(bd->max_queue, uarg);
896 	case SG_SET_COMMAND_Q: {
897 		int queue;
898 
899 		if (get_user(queue, uarg))
900 			return -EFAULT;
901 		if (queue < 1)
902 			return -EINVAL;
903 
904 		spin_lock_irq(&bd->lock);
905 		bd->max_queue = queue;
906 		spin_unlock_irq(&bd->lock);
907 		return 0;
908 	}
909 
910 	/*
911 	 * SCSI/sg ioctls
912 	 */
913 	case SG_GET_VERSION_NUM:
914 	case SCSI_IOCTL_GET_IDLUN:
915 	case SCSI_IOCTL_GET_BUS_NUMBER:
916 	case SG_SET_TIMEOUT:
917 	case SG_GET_TIMEOUT:
918 	case SG_GET_RESERVED_SIZE:
919 	case SG_SET_RESERVED_SIZE:
920 	case SG_EMULATED_HOST:
921 	case SCSI_IOCTL_SEND_COMMAND: {
922 		void __user *uarg = (void __user *) arg;
923 		return scsi_cmd_ioctl(bd->queue, NULL, file->f_mode, cmd, uarg);
924 	}
925 	case SG_IO: {
926 		struct request *rq;
927 		struct bio *bio, *bidi_bio = NULL;
928 		struct sg_io_v4 hdr;
929 		int at_head;
930 		u8 sense[SCSI_SENSE_BUFFERSIZE];
931 
932 		if (copy_from_user(&hdr, uarg, sizeof(hdr)))
933 			return -EFAULT;
934 
935 		rq = bsg_map_hdr(bd, &hdr, file->f_mode & FMODE_WRITE, sense);
936 		if (IS_ERR(rq))
937 			return PTR_ERR(rq);
938 
939 		bio = rq->bio;
940 		if (rq->next_rq)
941 			bidi_bio = rq->next_rq->bio;
942 
943 		at_head = (0 == (hdr.flags & BSG_FLAG_Q_AT_TAIL));
944 		blk_execute_rq(bd->queue, NULL, rq, at_head);
945 		ret = blk_complete_sgv4_hdr_rq(rq, &hdr, bio, bidi_bio);
946 
947 		if (copy_to_user(uarg, &hdr, sizeof(hdr)))
948 			return -EFAULT;
949 
950 		return ret;
951 	}
952 	/*
953 	 * block device ioctls
954 	 */
955 	default:
956 #if 0
957 		return ioctl_by_bdev(bd->bdev, cmd, arg);
958 #else
959 		return -ENOTTY;
960 #endif
961 	}
962 }
963 
964 static const struct file_operations bsg_fops = {
965 	.read		=	bsg_read,
966 	.write		=	bsg_write,
967 	.poll		=	bsg_poll,
968 	.open		=	bsg_open,
969 	.release	=	bsg_release,
970 	.unlocked_ioctl	=	bsg_ioctl,
971 	.owner		=	THIS_MODULE,
972 };
973 
974 void bsg_unregister_queue(struct request_queue *q)
975 {
976 	struct bsg_class_device *bcd = &q->bsg_dev;
977 
978 	if (!bcd->class_dev)
979 		return;
980 
981 	mutex_lock(&bsg_mutex);
982 	idr_remove(&bsg_minor_idr, bcd->minor);
983 	sysfs_remove_link(&q->kobj, "bsg");
984 	device_unregister(bcd->class_dev);
985 	bcd->class_dev = NULL;
986 	kref_put(&bcd->ref, bsg_kref_release_function);
987 	mutex_unlock(&bsg_mutex);
988 }
989 EXPORT_SYMBOL_GPL(bsg_unregister_queue);
990 
991 int bsg_register_queue(struct request_queue *q, struct device *parent,
992 		       const char *name, void (*release)(struct device *))
993 {
994 	struct bsg_class_device *bcd;
995 	dev_t dev;
996 	int ret, minor;
997 	struct device *class_dev = NULL;
998 	const char *devname;
999 
1000 	if (name)
1001 		devname = name;
1002 	else
1003 		devname = dev_name(parent);
1004 
1005 	/*
1006 	 * we need a proper transport to send commands, not a stacked device
1007 	 */
1008 	if (!q->request_fn)
1009 		return 0;
1010 
1011 	bcd = &q->bsg_dev;
1012 	memset(bcd, 0, sizeof(*bcd));
1013 
1014 	mutex_lock(&bsg_mutex);
1015 
1016 	ret = idr_pre_get(&bsg_minor_idr, GFP_KERNEL);
1017 	if (!ret) {
1018 		ret = -ENOMEM;
1019 		goto unlock;
1020 	}
1021 
1022 	ret = idr_get_new(&bsg_minor_idr, bcd, &minor);
1023 	if (ret < 0)
1024 		goto unlock;
1025 
1026 	if (minor >= BSG_MAX_DEVS) {
1027 		printk(KERN_ERR "bsg: too many bsg devices\n");
1028 		ret = -EINVAL;
1029 		goto remove_idr;
1030 	}
1031 
1032 	bcd->minor = minor;
1033 	bcd->queue = q;
1034 	bcd->parent = get_device(parent);
1035 	bcd->release = release;
1036 	kref_init(&bcd->ref);
1037 	dev = MKDEV(bsg_major, bcd->minor);
1038 	class_dev = device_create(bsg_class, parent, dev, NULL, "%s", devname);
1039 	if (IS_ERR(class_dev)) {
1040 		ret = PTR_ERR(class_dev);
1041 		goto put_dev;
1042 	}
1043 	bcd->class_dev = class_dev;
1044 
1045 	if (q->kobj.sd) {
1046 		ret = sysfs_create_link(&q->kobj, &bcd->class_dev->kobj, "bsg");
1047 		if (ret)
1048 			goto unregister_class_dev;
1049 	}
1050 
1051 	mutex_unlock(&bsg_mutex);
1052 	return 0;
1053 
1054 unregister_class_dev:
1055 	device_unregister(class_dev);
1056 put_dev:
1057 	put_device(parent);
1058 remove_idr:
1059 	idr_remove(&bsg_minor_idr, minor);
1060 unlock:
1061 	mutex_unlock(&bsg_mutex);
1062 	return ret;
1063 }
1064 EXPORT_SYMBOL_GPL(bsg_register_queue);
1065 
1066 static struct cdev bsg_cdev;
1067 
1068 static char *bsg_nodename(struct device *dev)
1069 {
1070 	return kasprintf(GFP_KERNEL, "bsg/%s", dev_name(dev));
1071 }
1072 
1073 static int __init bsg_init(void)
1074 {
1075 	int ret, i;
1076 	dev_t devid;
1077 
1078 	bsg_cmd_cachep = kmem_cache_create("bsg_cmd",
1079 				sizeof(struct bsg_command), 0, 0, NULL);
1080 	if (!bsg_cmd_cachep) {
1081 		printk(KERN_ERR "bsg: failed creating slab cache\n");
1082 		return -ENOMEM;
1083 	}
1084 
1085 	for (i = 0; i < BSG_LIST_ARRAY_SIZE; i++)
1086 		INIT_HLIST_HEAD(&bsg_device_list[i]);
1087 
1088 	bsg_class = class_create(THIS_MODULE, "bsg");
1089 	if (IS_ERR(bsg_class)) {
1090 		ret = PTR_ERR(bsg_class);
1091 		goto destroy_kmemcache;
1092 	}
1093 	bsg_class->nodename = bsg_nodename;
1094 
1095 	ret = alloc_chrdev_region(&devid, 0, BSG_MAX_DEVS, "bsg");
1096 	if (ret)
1097 		goto destroy_bsg_class;
1098 
1099 	bsg_major = MAJOR(devid);
1100 
1101 	cdev_init(&bsg_cdev, &bsg_fops);
1102 	ret = cdev_add(&bsg_cdev, MKDEV(bsg_major, 0), BSG_MAX_DEVS);
1103 	if (ret)
1104 		goto unregister_chrdev;
1105 
1106 	printk(KERN_INFO BSG_DESCRIPTION " version " BSG_VERSION
1107 	       " loaded (major %d)\n", bsg_major);
1108 	return 0;
1109 unregister_chrdev:
1110 	unregister_chrdev_region(MKDEV(bsg_major, 0), BSG_MAX_DEVS);
1111 destroy_bsg_class:
1112 	class_destroy(bsg_class);
1113 destroy_kmemcache:
1114 	kmem_cache_destroy(bsg_cmd_cachep);
1115 	return ret;
1116 }
1117 
1118 MODULE_AUTHOR("Jens Axboe");
1119 MODULE_DESCRIPTION(BSG_DESCRIPTION);
1120 MODULE_LICENSE("GPL");
1121 
1122 device_initcall(bsg_init);
1123