xref: /freebsd/sys/cam/ctl/ctl_backend_block.c (revision 3bdf775801b218aa5a89564839405b122f4b233e)
1 /*-
2  * Copyright (c) 2003 Silicon Graphics International Corp.
3  * Copyright (c) 2009-2011 Spectra Logic Corporation
4  * Copyright (c) 2012 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * Portions of this software were developed by Edward Tomasz Napierala
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions, and the following disclaimer,
15  *    without modification.
16  * 2. Redistributions in binary form must reproduce at minimum a disclaimer
17  *    substantially similar to the "NO WARRANTY" disclaimer below
18  *    ("Disclaimer") and any redistribution must be conditioned upon
19  *    including a substantially similar Disclaimer requirement for further
20  *    binary redistribution.
21  *
22  * NO WARRANTY
23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTIBILITY AND FITNESS FOR
26  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
27  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
31  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
32  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
33  * POSSIBILITY OF SUCH DAMAGES.
34  *
35  * $Id: //depot/users/kenm/FreeBSD-test2/sys/cam/ctl/ctl_backend_block.c#5 $
36  */
37 /*
38  * CAM Target Layer driver backend for block devices.
39  *
40  * Author: Ken Merry <ken@FreeBSD.org>
41  */
42 #include <sys/cdefs.h>
43 __FBSDID("$FreeBSD$");
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/types.h>
49 #include <sys/kthread.h>
50 #include <sys/bio.h>
51 #include <sys/fcntl.h>
52 #include <sys/limits.h>
53 #include <sys/lock.h>
54 #include <sys/mutex.h>
55 #include <sys/condvar.h>
56 #include <sys/malloc.h>
57 #include <sys/conf.h>
58 #include <sys/ioccom.h>
59 #include <sys/queue.h>
60 #include <sys/sbuf.h>
61 #include <sys/endian.h>
62 #include <sys/uio.h>
63 #include <sys/buf.h>
64 #include <sys/taskqueue.h>
65 #include <sys/vnode.h>
66 #include <sys/namei.h>
67 #include <sys/mount.h>
68 #include <sys/disk.h>
69 #include <sys/fcntl.h>
70 #include <sys/filedesc.h>
71 #include <sys/proc.h>
72 #include <sys/pcpu.h>
73 #include <sys/module.h>
74 #include <sys/sdt.h>
75 #include <sys/devicestat.h>
76 #include <sys/sysctl.h>
77 
78 #include <geom/geom.h>
79 
80 #include <cam/cam.h>
81 #include <cam/scsi/scsi_all.h>
82 #include <cam/scsi/scsi_da.h>
83 #include <cam/ctl/ctl_io.h>
84 #include <cam/ctl/ctl.h>
85 #include <cam/ctl/ctl_backend.h>
86 #include <cam/ctl/ctl_frontend_internal.h>
87 #include <cam/ctl/ctl_ioctl.h>
88 #include <cam/ctl/ctl_scsi_all.h>
89 #include <cam/ctl/ctl_error.h>
90 
91 /*
92  * The idea here is that we'll allocate enough S/G space to hold a 16MB
93  * I/O.  If we get an I/O larger than that, we'll reject it.
94  */
95 #define	CTLBLK_MAX_IO_SIZE	(16 * 1024 * 1024)
96 #define	CTLBLK_MAX_SEGS		(CTLBLK_MAX_IO_SIZE / MAXPHYS) + 1
97 
98 #ifdef CTLBLK_DEBUG
99 #define DPRINTF(fmt, args...) \
100     printf("cbb(%s:%d): " fmt, __FUNCTION__, __LINE__, ##args)
101 #else
102 #define DPRINTF(fmt, args...) do {} while(0)
103 #endif
104 
105 SDT_PROVIDER_DEFINE(cbb);
106 
107 typedef enum {
108 	CTL_BE_BLOCK_LUN_UNCONFIGURED	= 0x01,
109 	CTL_BE_BLOCK_LUN_CONFIG_ERR	= 0x02,
110 	CTL_BE_BLOCK_LUN_WAITING	= 0x04,
111 	CTL_BE_BLOCK_LUN_MULTI_THREAD	= 0x08
112 } ctl_be_block_lun_flags;
113 
114 typedef enum {
115 	CTL_BE_BLOCK_NONE,
116 	CTL_BE_BLOCK_DEV,
117 	CTL_BE_BLOCK_FILE
118 } ctl_be_block_type;
119 
120 struct ctl_be_block_devdata {
121 	struct cdev *cdev;
122 	struct cdevsw *csw;
123 	int dev_ref;
124 };
125 
126 struct ctl_be_block_filedata {
127 	struct ucred *cred;
128 };
129 
130 union ctl_be_block_bedata {
131 	struct ctl_be_block_devdata dev;
132 	struct ctl_be_block_filedata file;
133 };
134 
135 struct ctl_be_block_io;
136 struct ctl_be_block_lun;
137 
138 typedef void (*cbb_dispatch_t)(struct ctl_be_block_lun *be_lun,
139 			       struct ctl_be_block_io *beio);
140 
141 /*
142  * Backend LUN structure.  There is a 1:1 mapping between a block device
143  * and a backend block LUN, and between a backend block LUN and a CTL LUN.
144  */
145 struct ctl_be_block_lun {
146 	struct ctl_block_disk *disk;
147 	char lunname[32];
148 	char *dev_path;
149 	ctl_be_block_type dev_type;
150 	struct vnode *vn;
151 	union ctl_be_block_bedata backend;
152 	cbb_dispatch_t dispatch;
153 	cbb_dispatch_t lun_flush;
154 	cbb_dispatch_t unmap;
155 	struct mtx lock;
156 	uma_zone_t lun_zone;
157 	uint64_t size_blocks;
158 	uint64_t size_bytes;
159 	uint32_t blocksize;
160 	int blocksize_shift;
161 	uint16_t pblockexp;
162 	uint16_t pblockoff;
163 	struct ctl_be_block_softc *softc;
164 	struct devstat *disk_stats;
165 	ctl_be_block_lun_flags flags;
166 	STAILQ_ENTRY(ctl_be_block_lun) links;
167 	struct ctl_be_lun ctl_be_lun;
168 	struct taskqueue *io_taskqueue;
169 	struct task io_task;
170 	int num_threads;
171 	STAILQ_HEAD(, ctl_io_hdr) input_queue;
172 	STAILQ_HEAD(, ctl_io_hdr) config_write_queue;
173 	STAILQ_HEAD(, ctl_io_hdr) datamove_queue;
174 };
175 
176 /*
177  * Overall softc structure for the block backend module.
178  */
179 struct ctl_be_block_softc {
180 	struct mtx			 lock;
181 	int				 num_disks;
182 	STAILQ_HEAD(, ctl_block_disk)	 disk_list;
183 	int				 num_luns;
184 	STAILQ_HEAD(, ctl_be_block_lun)	 lun_list;
185 };
186 
187 static struct ctl_be_block_softc backend_block_softc;
188 
189 /*
190  * Per-I/O information.
191  */
192 struct ctl_be_block_io {
193 	union ctl_io			*io;
194 	struct ctl_sg_entry		sg_segs[CTLBLK_MAX_SEGS];
195 	struct iovec			xiovecs[CTLBLK_MAX_SEGS];
196 	int				bio_cmd;
197 	int				bio_flags;
198 	int				num_segs;
199 	int				num_bios_sent;
200 	int				num_bios_done;
201 	int				send_complete;
202 	int				num_errors;
203 	struct bintime			ds_t0;
204 	devstat_tag_type		ds_tag_type;
205 	devstat_trans_flags		ds_trans_type;
206 	uint64_t			io_len;
207 	uint64_t			io_offset;
208 	struct ctl_be_block_softc	*softc;
209 	struct ctl_be_block_lun		*lun;
210 	void (*beio_cont)(struct ctl_be_block_io *beio); /* to continue processing */
211 };
212 
213 static int cbb_num_threads = 14;
214 TUNABLE_INT("kern.cam.ctl.block.num_threads", &cbb_num_threads);
215 SYSCTL_NODE(_kern_cam_ctl, OID_AUTO, block, CTLFLAG_RD, 0,
216 	    "CAM Target Layer Block Backend");
217 SYSCTL_INT(_kern_cam_ctl_block, OID_AUTO, num_threads, CTLFLAG_RW,
218            &cbb_num_threads, 0, "Number of threads per backing file");
219 
220 static struct ctl_be_block_io *ctl_alloc_beio(struct ctl_be_block_softc *softc);
221 static void ctl_free_beio(struct ctl_be_block_io *beio);
222 static void ctl_complete_beio(struct ctl_be_block_io *beio);
223 static int ctl_be_block_move_done(union ctl_io *io);
224 static void ctl_be_block_biodone(struct bio *bio);
225 static void ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
226 				    struct ctl_be_block_io *beio);
227 static void ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
228 				       struct ctl_be_block_io *beio);
229 static void ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
230 				   struct ctl_be_block_io *beio);
231 static void ctl_be_block_unmap_dev(struct ctl_be_block_lun *be_lun,
232 				   struct ctl_be_block_io *beio);
233 static void ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
234 				      struct ctl_be_block_io *beio);
235 static void ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
236 				    union ctl_io *io);
237 static void ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
238 				  union ctl_io *io);
239 static void ctl_be_block_worker(void *context, int pending);
240 static int ctl_be_block_submit(union ctl_io *io);
241 static int ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
242 				   int flag, struct thread *td);
243 static int ctl_be_block_open_file(struct ctl_be_block_lun *be_lun,
244 				  struct ctl_lun_req *req);
245 static int ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun,
246 				 struct ctl_lun_req *req);
247 static int ctl_be_block_close(struct ctl_be_block_lun *be_lun);
248 static int ctl_be_block_open(struct ctl_be_block_softc *softc,
249 			     struct ctl_be_block_lun *be_lun,
250 			     struct ctl_lun_req *req);
251 static int ctl_be_block_create(struct ctl_be_block_softc *softc,
252 			       struct ctl_lun_req *req);
253 static int ctl_be_block_rm(struct ctl_be_block_softc *softc,
254 			   struct ctl_lun_req *req);
255 static int ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
256 				  struct ctl_lun_req *req);
257 static int ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
258 				 struct ctl_lun_req *req);
259 static int ctl_be_block_modify(struct ctl_be_block_softc *softc,
260 			   struct ctl_lun_req *req);
261 static void ctl_be_block_lun_shutdown(void *be_lun);
262 static void ctl_be_block_lun_config_status(void *be_lun,
263 					   ctl_lun_config_status status);
264 static int ctl_be_block_config_write(union ctl_io *io);
265 static int ctl_be_block_config_read(union ctl_io *io);
266 static int ctl_be_block_lun_info(void *be_lun, struct sbuf *sb);
267 int ctl_be_block_init(void);
268 
269 static struct ctl_backend_driver ctl_be_block_driver =
270 {
271 	.name = "block",
272 	.flags = CTL_BE_FLAG_HAS_CONFIG,
273 	.init = ctl_be_block_init,
274 	.data_submit = ctl_be_block_submit,
275 	.data_move_done = ctl_be_block_move_done,
276 	.config_read = ctl_be_block_config_read,
277 	.config_write = ctl_be_block_config_write,
278 	.ioctl = ctl_be_block_ioctl,
279 	.lun_info = ctl_be_block_lun_info
280 };
281 
282 MALLOC_DEFINE(M_CTLBLK, "ctlblk", "Memory used for CTL block backend");
283 CTL_BACKEND_DECLARE(cbb, ctl_be_block_driver);
284 
285 static uma_zone_t beio_zone;
286 
287 static struct ctl_be_block_io *
288 ctl_alloc_beio(struct ctl_be_block_softc *softc)
289 {
290 	struct ctl_be_block_io *beio;
291 
292 	beio = uma_zalloc(beio_zone, M_WAITOK | M_ZERO);
293 	beio->softc = softc;
294 	return (beio);
295 }
296 
297 static void
298 ctl_free_beio(struct ctl_be_block_io *beio)
299 {
300 	int duplicate_free;
301 	int i;
302 
303 	duplicate_free = 0;
304 
305 	for (i = 0; i < beio->num_segs; i++) {
306 		if (beio->sg_segs[i].addr == NULL)
307 			duplicate_free++;
308 
309 		uma_zfree(beio->lun->lun_zone, beio->sg_segs[i].addr);
310 		beio->sg_segs[i].addr = NULL;
311 	}
312 
313 	if (duplicate_free > 0) {
314 		printf("%s: %d duplicate frees out of %d segments\n", __func__,
315 		       duplicate_free, beio->num_segs);
316 	}
317 
318 	uma_zfree(beio_zone, beio);
319 }
320 
321 static void
322 ctl_complete_beio(struct ctl_be_block_io *beio)
323 {
324 	union ctl_io *io;
325 	int io_len;
326 
327 	io = beio->io;
328 
329 	if ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_SUCCESS)
330 		io_len = beio->io_len;
331 	else
332 		io_len = 0;
333 
334 	devstat_end_transaction(beio->lun->disk_stats,
335 				/*bytes*/ io_len,
336 				beio->ds_tag_type,
337 				beio->ds_trans_type,
338 				/*now*/ NULL,
339 				/*then*/&beio->ds_t0);
340 
341 	if (beio->beio_cont != NULL) {
342 		beio->beio_cont(beio);
343 	} else {
344 		ctl_free_beio(beio);
345 		ctl_done(io);
346 	}
347 }
348 
349 static int
350 ctl_be_block_move_done(union ctl_io *io)
351 {
352 	struct ctl_be_block_io *beio;
353 	struct ctl_be_block_lun *be_lun;
354 #ifdef CTL_TIME_IO
355 	struct bintime cur_bt;
356 #endif
357 
358 	beio = (struct ctl_be_block_io *)
359 		io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr;
360 
361 	be_lun = beio->lun;
362 
363 	DPRINTF("entered\n");
364 
365 #ifdef CTL_TIME_IO
366 	getbintime(&cur_bt);
367 	bintime_sub(&cur_bt, &io->io_hdr.dma_start_bt);
368 	bintime_add(&io->io_hdr.dma_bt, &cur_bt);
369 	io->io_hdr.num_dmas++;
370 #endif
371 
372 	/*
373 	 * We set status at this point for read commands, and write
374 	 * commands with errors.
375 	 */
376 	if ((beio->bio_cmd == BIO_READ)
377 	 && (io->io_hdr.port_status == 0)
378 	 && ((io->io_hdr.flags & CTL_FLAG_ABORT) == 0)
379 	 && ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_STATUS_NONE))
380 		ctl_set_success(&io->scsiio);
381 	else if ((io->io_hdr.port_status != 0)
382 	      && ((io->io_hdr.flags & CTL_FLAG_ABORT) == 0)
383 	      && ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_STATUS_NONE)) {
384 		/*
385 		 * For hardware error sense keys, the sense key
386 		 * specific value is defined to be a retry count,
387 		 * but we use it to pass back an internal FETD
388 		 * error code.  XXX KDM  Hopefully the FETD is only
389 		 * using 16 bits for an error code, since that's
390 		 * all the space we have in the sks field.
391 		 */
392 		ctl_set_internal_failure(&io->scsiio,
393 					 /*sks_valid*/ 1,
394 					 /*retry_count*/
395 					 io->io_hdr.port_status);
396 	}
397 
398 	/*
399 	 * If this is a read, or a write with errors, it is done.
400 	 */
401 	if ((beio->bio_cmd == BIO_READ)
402 	 || ((io->io_hdr.flags & CTL_FLAG_ABORT) != 0)
403 	 || ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE)) {
404 		ctl_complete_beio(beio);
405 		return (0);
406 	}
407 
408 	/*
409 	 * At this point, we have a write and the DMA completed
410 	 * successfully.  We now have to queue it to the task queue to
411 	 * execute the backend I/O.  That is because we do blocking
412 	 * memory allocations, and in the file backing case, blocking I/O.
413 	 * This move done routine is generally called in the SIM's
414 	 * interrupt context, and therefore we cannot block.
415 	 */
416 	mtx_lock(&be_lun->lock);
417 	/*
418 	 * XXX KDM make sure that links is okay to use at this point.
419 	 * Otherwise, we either need to add another field to ctl_io_hdr,
420 	 * or deal with resource allocation here.
421 	 */
422 	STAILQ_INSERT_TAIL(&be_lun->datamove_queue, &io->io_hdr, links);
423 	mtx_unlock(&be_lun->lock);
424 
425 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
426 
427 	return (0);
428 }
429 
430 static void
431 ctl_be_block_biodone(struct bio *bio)
432 {
433 	struct ctl_be_block_io *beio;
434 	struct ctl_be_block_lun *be_lun;
435 	union ctl_io *io;
436 	int error;
437 
438 	beio = bio->bio_caller1;
439 	be_lun = beio->lun;
440 	io = beio->io;
441 
442 	DPRINTF("entered\n");
443 
444 	error = bio->bio_error;
445 	mtx_lock(&be_lun->lock);
446 	if (error != 0)
447 		beio->num_errors++;
448 
449 	beio->num_bios_done++;
450 
451 	/*
452 	 * XXX KDM will this cause WITNESS to complain?  Holding a lock
453 	 * during the free might cause it to complain.
454 	 */
455 	g_destroy_bio(bio);
456 
457 	/*
458 	 * If the send complete bit isn't set, or we aren't the last I/O to
459 	 * complete, then we're done.
460 	 */
461 	if ((beio->send_complete == 0)
462 	 || (beio->num_bios_done < beio->num_bios_sent)) {
463 		mtx_unlock(&be_lun->lock);
464 		return;
465 	}
466 
467 	/*
468 	 * At this point, we've verified that we are the last I/O to
469 	 * complete, so it's safe to drop the lock.
470 	 */
471 	mtx_unlock(&be_lun->lock);
472 
473 	/*
474 	 * If there are any errors from the backing device, we fail the
475 	 * entire I/O with a medium error.
476 	 */
477 	if (beio->num_errors > 0) {
478 		if (error == EOPNOTSUPP) {
479 			ctl_set_invalid_opcode(&io->scsiio);
480 		} else if (beio->bio_cmd == BIO_FLUSH) {
481 			/* XXX KDM is there is a better error here? */
482 			ctl_set_internal_failure(&io->scsiio,
483 						 /*sks_valid*/ 1,
484 						 /*retry_count*/ 0xbad2);
485 		} else
486 			ctl_set_medium_error(&io->scsiio);
487 		ctl_complete_beio(beio);
488 		return;
489 	}
490 
491 	/*
492 	 * If this is a write, a flush or a delete, we're all done.
493 	 * If this is a read, we can now send the data to the user.
494 	 */
495 	if ((beio->bio_cmd == BIO_WRITE)
496 	 || (beio->bio_cmd == BIO_FLUSH)
497 	 || (beio->bio_cmd == BIO_DELETE)) {
498 		ctl_set_success(&io->scsiio);
499 		ctl_complete_beio(beio);
500 	} else {
501 		io->scsiio.be_move_done = ctl_be_block_move_done;
502 		io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
503 		io->scsiio.kern_data_len = beio->io_len;
504 		io->scsiio.kern_total_len = beio->io_len;
505 		io->scsiio.kern_rel_offset = 0;
506 		io->scsiio.kern_data_resid = 0;
507 		io->scsiio.kern_sg_entries = beio->num_segs;
508 		io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
509 #ifdef CTL_TIME_IO
510         	getbintime(&io->io_hdr.dma_start_bt);
511 #endif
512 		ctl_datamove(io);
513 	}
514 }
515 
516 static void
517 ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
518 			struct ctl_be_block_io *beio)
519 {
520 	union ctl_io *io;
521 	struct mount *mountpoint;
522 	int error, lock_flags;
523 
524 	DPRINTF("entered\n");
525 
526 	io = beio->io;
527 
528        	(void) vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
529 
530 	if (MNT_SHARED_WRITES(mountpoint)
531 	 || ((mountpoint == NULL)
532 	  && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
533 		lock_flags = LK_SHARED;
534 	else
535 		lock_flags = LK_EXCLUSIVE;
536 
537 	vn_lock(be_lun->vn, lock_flags | LK_RETRY);
538 
539 	binuptime(&beio->ds_t0);
540 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
541 
542 	error = VOP_FSYNC(be_lun->vn, MNT_WAIT, curthread);
543 	VOP_UNLOCK(be_lun->vn, 0);
544 
545 	vn_finished_write(mountpoint);
546 
547 	if (error == 0)
548 		ctl_set_success(&io->scsiio);
549 	else {
550 		/* XXX KDM is there is a better error here? */
551 		ctl_set_internal_failure(&io->scsiio,
552 					 /*sks_valid*/ 1,
553 					 /*retry_count*/ 0xbad1);
554 	}
555 
556 	ctl_complete_beio(beio);
557 }
558 
559 SDT_PROBE_DEFINE1(cbb, kernel, read, file_start, "uint64_t");
560 SDT_PROBE_DEFINE1(cbb, kernel, write, file_start, "uint64_t");
561 SDT_PROBE_DEFINE1(cbb, kernel, read, file_done,"uint64_t");
562 SDT_PROBE_DEFINE1(cbb, kernel, write, file_done, "uint64_t");
563 
564 static void
565 ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
566 			   struct ctl_be_block_io *beio)
567 {
568 	struct ctl_be_block_filedata *file_data;
569 	union ctl_io *io;
570 	struct uio xuio;
571 	struct iovec *xiovec;
572 	int flags;
573 	int error, i;
574 
575 	DPRINTF("entered\n");
576 
577 	file_data = &be_lun->backend.file;
578 	io = beio->io;
579 	flags = beio->bio_flags;
580 
581 	if (beio->bio_cmd == BIO_READ) {
582 		SDT_PROBE(cbb, kernel, read, file_start, 0, 0, 0, 0, 0);
583 	} else {
584 		SDT_PROBE(cbb, kernel, write, file_start, 0, 0, 0, 0, 0);
585 	}
586 
587 	bzero(&xuio, sizeof(xuio));
588 	if (beio->bio_cmd == BIO_READ)
589 		xuio.uio_rw = UIO_READ;
590 	else
591 		xuio.uio_rw = UIO_WRITE;
592 
593 	xuio.uio_offset = beio->io_offset;
594 	xuio.uio_resid = beio->io_len;
595 	xuio.uio_segflg = UIO_SYSSPACE;
596 	xuio.uio_iov = beio->xiovecs;
597 	xuio.uio_iovcnt = beio->num_segs;
598 	xuio.uio_td = curthread;
599 
600 	for (i = 0, xiovec = xuio.uio_iov; i < xuio.uio_iovcnt; i++, xiovec++) {
601 		xiovec->iov_base = beio->sg_segs[i].addr;
602 		xiovec->iov_len = beio->sg_segs[i].len;
603 	}
604 
605 	if (beio->bio_cmd == BIO_READ) {
606 		vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
607 
608 		binuptime(&beio->ds_t0);
609 		devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
610 
611 		/*
612 		 * UFS pays attention to IO_DIRECT for reads.  If the
613 		 * DIRECTIO option is configured into the kernel, it calls
614 		 * ffs_rawread().  But that only works for single-segment
615 		 * uios with user space addresses.  In our case, with a
616 		 * kernel uio, it still reads into the buffer cache, but it
617 		 * will just try to release the buffer from the cache later
618 		 * on in ffs_read().
619 		 *
620 		 * ZFS does not pay attention to IO_DIRECT for reads.
621 		 *
622 		 * UFS does not pay attention to IO_SYNC for reads.
623 		 *
624 		 * ZFS pays attention to IO_SYNC (which translates into the
625 		 * Solaris define FRSYNC for zfs_read()) for reads.  It
626 		 * attempts to sync the file before reading.
627 		 *
628 		 * So, to attempt to provide some barrier semantics in the
629 		 * BIO_ORDERED case, set both IO_DIRECT and IO_SYNC.
630 		 */
631 		error = VOP_READ(be_lun->vn, &xuio, (flags & BIO_ORDERED) ?
632 				 (IO_DIRECT|IO_SYNC) : 0, file_data->cred);
633 
634 		VOP_UNLOCK(be_lun->vn, 0);
635 	} else {
636 		struct mount *mountpoint;
637 		int lock_flags;
638 
639 		(void)vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
640 
641 		if (MNT_SHARED_WRITES(mountpoint)
642 		 || ((mountpoint == NULL)
643 		  && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
644 			lock_flags = LK_SHARED;
645 		else
646 			lock_flags = LK_EXCLUSIVE;
647 
648 		vn_lock(be_lun->vn, lock_flags | LK_RETRY);
649 
650 		binuptime(&beio->ds_t0);
651 		devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
652 
653 		/*
654 		 * UFS pays attention to IO_DIRECT for writes.  The write
655 		 * is done asynchronously.  (Normally the write would just
656 		 * get put into cache.
657 		 *
658 		 * UFS pays attention to IO_SYNC for writes.  It will
659 		 * attempt to write the buffer out synchronously if that
660 		 * flag is set.
661 		 *
662 		 * ZFS does not pay attention to IO_DIRECT for writes.
663 		 *
664 		 * ZFS pays attention to IO_SYNC (a.k.a. FSYNC or FRSYNC)
665 		 * for writes.  It will flush the transaction from the
666 		 * cache before returning.
667 		 *
668 		 * So if we've got the BIO_ORDERED flag set, we want
669 		 * IO_SYNC in either the UFS or ZFS case.
670 		 */
671 		error = VOP_WRITE(be_lun->vn, &xuio, (flags & BIO_ORDERED) ?
672 				  IO_SYNC : 0, file_data->cred);
673 		VOP_UNLOCK(be_lun->vn, 0);
674 
675 		vn_finished_write(mountpoint);
676         }
677 
678 	/*
679 	 * If we got an error, set the sense data to "MEDIUM ERROR" and
680 	 * return the I/O to the user.
681 	 */
682 	if (error != 0) {
683 		char path_str[32];
684 
685 		ctl_scsi_path_string(io, path_str, sizeof(path_str));
686 		/*
687 		 * XXX KDM ZFS returns ENOSPC when the underlying
688 		 * filesystem fills up.  What kind of SCSI error should we
689 		 * return for that?
690 		 */
691 		printf("%s%s command returned errno %d\n", path_str,
692 		       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE", error);
693 		ctl_set_medium_error(&io->scsiio);
694 		ctl_complete_beio(beio);
695 		return;
696 	}
697 
698 	/*
699 	 * If this is a write, we're all done.
700 	 * If this is a read, we can now send the data to the user.
701 	 */
702 	if (beio->bio_cmd == BIO_WRITE) {
703 		ctl_set_success(&io->scsiio);
704 		SDT_PROBE(cbb, kernel, write, file_done, 0, 0, 0, 0, 0);
705 		ctl_complete_beio(beio);
706 	} else {
707 		SDT_PROBE(cbb, kernel, read, file_done, 0, 0, 0, 0, 0);
708 		io->scsiio.be_move_done = ctl_be_block_move_done;
709 		io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
710 		io->scsiio.kern_data_len = beio->io_len;
711 		io->scsiio.kern_total_len = beio->io_len;
712 		io->scsiio.kern_rel_offset = 0;
713 		io->scsiio.kern_data_resid = 0;
714 		io->scsiio.kern_sg_entries = beio->num_segs;
715 		io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
716 #ifdef CTL_TIME_IO
717         	getbintime(&io->io_hdr.dma_start_bt);
718 #endif
719 		ctl_datamove(io);
720 	}
721 }
722 
723 static void
724 ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
725 		       struct ctl_be_block_io *beio)
726 {
727 	struct bio *bio;
728 	union ctl_io *io;
729 	struct ctl_be_block_devdata *dev_data;
730 
731 	dev_data = &be_lun->backend.dev;
732 	io = beio->io;
733 
734 	DPRINTF("entered\n");
735 
736 	/* This can't fail, it's a blocking allocation. */
737 	bio = g_alloc_bio();
738 
739 	bio->bio_cmd	    = BIO_FLUSH;
740 	bio->bio_flags	   |= BIO_ORDERED;
741 	bio->bio_dev	    = dev_data->cdev;
742 	bio->bio_offset	    = 0;
743 	bio->bio_data	    = 0;
744 	bio->bio_done	    = ctl_be_block_biodone;
745 	bio->bio_caller1    = beio;
746 	bio->bio_pblkno	    = 0;
747 
748 	/*
749 	 * We don't need to acquire the LUN lock here, because we are only
750 	 * sending one bio, and so there is no other context to synchronize
751 	 * with.
752 	 */
753 	beio->num_bios_sent = 1;
754 	beio->send_complete = 1;
755 
756 	binuptime(&beio->ds_t0);
757 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
758 
759 	(*dev_data->csw->d_strategy)(bio);
760 }
761 
762 static void
763 ctl_be_block_unmap_dev_range(struct ctl_be_block_lun *be_lun,
764 		       struct ctl_be_block_io *beio,
765 		       uint64_t off, uint64_t len, int last)
766 {
767 	struct bio *bio;
768 	struct ctl_be_block_devdata *dev_data;
769 	uint64_t maxlen;
770 
771 	dev_data = &be_lun->backend.dev;
772 	maxlen = LONG_MAX - (LONG_MAX % be_lun->blocksize);
773 	while (len > 0) {
774 		bio = g_alloc_bio();
775 		bio->bio_cmd	    = BIO_DELETE;
776 		bio->bio_flags	   |= beio->bio_flags;
777 		bio->bio_dev	    = dev_data->cdev;
778 		bio->bio_offset	    = off;
779 		bio->bio_length	    = MIN(len, maxlen);
780 		bio->bio_data	    = 0;
781 		bio->bio_done	    = ctl_be_block_biodone;
782 		bio->bio_caller1    = beio;
783 		bio->bio_pblkno     = off / be_lun->blocksize;
784 
785 		off += bio->bio_length;
786 		len -= bio->bio_length;
787 
788 		mtx_lock(&be_lun->lock);
789 		beio->num_bios_sent++;
790 		if (last && len == 0)
791 			beio->send_complete = 1;
792 		mtx_unlock(&be_lun->lock);
793 
794 		(*dev_data->csw->d_strategy)(bio);
795 	}
796 }
797 
798 static void
799 ctl_be_block_unmap_dev(struct ctl_be_block_lun *be_lun,
800 		       struct ctl_be_block_io *beio)
801 {
802 	union ctl_io *io;
803 	struct ctl_be_block_devdata *dev_data;
804 	struct ctl_ptr_len_flags ptrlen;
805 	struct scsi_unmap_desc *buf, *end;
806 	uint64_t len;
807 
808 	dev_data = &be_lun->backend.dev;
809 	io = beio->io;
810 
811 	DPRINTF("entered\n");
812 
813 	binuptime(&beio->ds_t0);
814 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
815 
816 	if (beio->io_offset == -1) {
817 		beio->io_len = 0;
818 		memcpy(&ptrlen, io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN].bytes,
819 		       sizeof(ptrlen));
820 		buf = (struct scsi_unmap_desc *)ptrlen.ptr;
821 		end = buf + ptrlen.len / sizeof(*buf);
822 		for (; buf < end; buf++) {
823 			len = (uint64_t)scsi_4btoul(buf->length) *
824 			    be_lun->blocksize;
825 			beio->io_len += len;
826 			ctl_be_block_unmap_dev_range(be_lun, beio,
827 			    scsi_8btou64(buf->lba) * be_lun->blocksize, len,
828 			    (end - buf < 2) ? TRUE : FALSE);
829 		}
830 	} else
831 		ctl_be_block_unmap_dev_range(be_lun, beio,
832 		    beio->io_offset, beio->io_len, TRUE);
833 }
834 
835 static void
836 ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
837 			  struct ctl_be_block_io *beio)
838 {
839 	int i;
840 	struct bio *bio;
841 	struct ctl_be_block_devdata *dev_data;
842 	off_t cur_offset;
843 	int max_iosize;
844 
845 	DPRINTF("entered\n");
846 
847 	dev_data = &be_lun->backend.dev;
848 
849 	/*
850 	 * We have to limit our I/O size to the maximum supported by the
851 	 * backend device.  Hopefully it is MAXPHYS.  If the driver doesn't
852 	 * set it properly, use DFLTPHYS.
853 	 */
854 	max_iosize = dev_data->cdev->si_iosize_max;
855 	if (max_iosize < PAGE_SIZE)
856 		max_iosize = DFLTPHYS;
857 
858 	cur_offset = beio->io_offset;
859 
860 	/*
861 	 * XXX KDM need to accurately reflect the number of I/Os outstanding
862 	 * to a device.
863 	 */
864 	binuptime(&beio->ds_t0);
865 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
866 
867 	for (i = 0; i < beio->num_segs; i++) {
868 		size_t cur_size;
869 		uint8_t *cur_ptr;
870 
871 		cur_size = beio->sg_segs[i].len;
872 		cur_ptr = beio->sg_segs[i].addr;
873 
874 		while (cur_size > 0) {
875 			/* This can't fail, it's a blocking allocation. */
876 			bio = g_alloc_bio();
877 
878 			KASSERT(bio != NULL, ("g_alloc_bio() failed!\n"));
879 
880 			bio->bio_cmd = beio->bio_cmd;
881 			bio->bio_flags |= beio->bio_flags;
882 			bio->bio_dev = dev_data->cdev;
883 			bio->bio_caller1 = beio;
884 			bio->bio_length = min(cur_size, max_iosize);
885 			bio->bio_offset = cur_offset;
886 			bio->bio_data = cur_ptr;
887 			bio->bio_done = ctl_be_block_biodone;
888 			bio->bio_pblkno = cur_offset / be_lun->blocksize;
889 
890 			cur_offset += bio->bio_length;
891 			cur_ptr += bio->bio_length;
892 			cur_size -= bio->bio_length;
893 
894 			/*
895 			 * Make sure we set the complete bit just before we
896 			 * issue the last bio so we don't wind up with a
897 			 * race.
898 			 *
899 			 * Use the LUN mutex here instead of a combination
900 			 * of atomic variables for simplicity.
901 			 *
902 			 * XXX KDM we could have a per-IO lock, but that
903 			 * would cause additional per-IO setup and teardown
904 			 * overhead.  Hopefully there won't be too much
905 			 * contention on the LUN lock.
906 			 */
907 			mtx_lock(&be_lun->lock);
908 
909 			beio->num_bios_sent++;
910 
911 			if ((i == beio->num_segs - 1)
912 			 && (cur_size == 0))
913 				beio->send_complete = 1;
914 
915 			mtx_unlock(&be_lun->lock);
916 
917 			(*dev_data->csw->d_strategy)(bio);
918 		}
919 	}
920 }
921 
922 static void
923 ctl_be_block_cw_done_ws(struct ctl_be_block_io *beio)
924 {
925 	union ctl_io *io;
926 
927 	io = beio->io;
928 	ctl_free_beio(beio);
929 	if (((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE)
930 	  && ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_SUCCESS)) {
931 		ctl_config_write_done(io);
932 		return;
933 	}
934 
935 	ctl_be_block_config_write(io);
936 }
937 
938 static void
939 ctl_be_block_cw_dispatch_ws(struct ctl_be_block_lun *be_lun,
940 			    union ctl_io *io)
941 {
942 	struct ctl_be_block_io *beio;
943 	struct ctl_be_block_softc *softc;
944 	struct ctl_lba_len_flags lbalen;
945 	uint64_t len_left, lba;
946 	int i, seglen;
947 	uint8_t *buf, *end;
948 
949 	DPRINTF("entered\n");
950 
951 	beio = io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr;
952 	softc = be_lun->softc;
953 	memcpy(&lbalen, io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN].bytes,
954 	       sizeof(lbalen));
955 
956 	if (lbalen.flags & ~(SWS_LBDATA | SWS_UNMAP) ||
957 	    (lbalen.flags & SWS_UNMAP && be_lun->unmap == NULL)) {
958 		ctl_free_beio(beio);
959 		ctl_set_invalid_field(&io->scsiio,
960 				      /*sks_valid*/ 1,
961 				      /*command*/ 1,
962 				      /*field*/ 1,
963 				      /*bit_valid*/ 0,
964 				      /*bit*/ 0);
965 		ctl_config_write_done(io);
966 		return;
967 	}
968 
969 	/*
970 	 * If the I/O came down with an ordered or head of queue tag, set
971 	 * the BIO_ORDERED attribute.  For head of queue tags, that's
972 	 * pretty much the best we can do.
973 	 */
974 	if ((io->scsiio.tag_type == CTL_TAG_ORDERED)
975 	 || (io->scsiio.tag_type == CTL_TAG_HEAD_OF_QUEUE))
976 		beio->bio_flags = BIO_ORDERED;
977 
978 	switch (io->scsiio.tag_type) {
979 	case CTL_TAG_ORDERED:
980 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
981 		break;
982 	case CTL_TAG_HEAD_OF_QUEUE:
983 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
984 		break;
985 	case CTL_TAG_UNTAGGED:
986 	case CTL_TAG_SIMPLE:
987 	case CTL_TAG_ACA:
988 	default:
989 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
990 		break;
991 	}
992 
993 	if (lbalen.flags & SWS_UNMAP) {
994 		beio->io_offset = lbalen.lba * be_lun->blocksize;
995 		beio->io_len = (uint64_t)lbalen.len * be_lun->blocksize;
996 		beio->bio_cmd = BIO_DELETE;
997 		beio->ds_trans_type = DEVSTAT_FREE;
998 
999 		be_lun->unmap(be_lun, beio);
1000 		return;
1001 	}
1002 
1003 	beio->bio_cmd = BIO_WRITE;
1004 	beio->ds_trans_type = DEVSTAT_WRITE;
1005 
1006 	DPRINTF("WRITE SAME at LBA %jx len %u\n",
1007 	       (uintmax_t)lbalen.lba, lbalen.len);
1008 
1009 	len_left = (uint64_t)lbalen.len * be_lun->blocksize;
1010 	for (i = 0, lba = 0; i < CTLBLK_MAX_SEGS && len_left > 0; i++) {
1011 
1012 		/*
1013 		 * Setup the S/G entry for this chunk.
1014 		 */
1015 		seglen = MIN(MAXPHYS, len_left);
1016 		seglen -= seglen % be_lun->blocksize;
1017 		beio->sg_segs[i].len = seglen;
1018 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1019 
1020 		DPRINTF("segment %d addr %p len %zd\n", i,
1021 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1022 
1023 		beio->num_segs++;
1024 		len_left -= seglen;
1025 
1026 		buf = beio->sg_segs[i].addr;
1027 		end = buf + seglen;
1028 		for (; buf < end; buf += be_lun->blocksize) {
1029 			memcpy(buf, io->scsiio.kern_data_ptr, be_lun->blocksize);
1030 			if (lbalen.flags & SWS_LBDATA)
1031 				scsi_ulto4b(lbalen.lba + lba, buf);
1032 			lba++;
1033 		}
1034 	}
1035 
1036 	beio->io_offset = lbalen.lba * be_lun->blocksize;
1037 	beio->io_len = lba * be_lun->blocksize;
1038 
1039 	/* We can not do all in one run. Correct and schedule rerun. */
1040 	if (len_left > 0) {
1041 		lbalen.lba += lba;
1042 		lbalen.len -= lba;
1043 		memcpy(io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN].bytes, &lbalen,
1044 		       sizeof(lbalen));
1045 		beio->beio_cont = ctl_be_block_cw_done_ws;
1046 	}
1047 
1048 	be_lun->dispatch(be_lun, beio);
1049 }
1050 
1051 static void
1052 ctl_be_block_cw_dispatch_unmap(struct ctl_be_block_lun *be_lun,
1053 			    union ctl_io *io)
1054 {
1055 	struct ctl_be_block_io *beio;
1056 	struct ctl_be_block_softc *softc;
1057 	struct ctl_ptr_len_flags ptrlen;
1058 
1059 	DPRINTF("entered\n");
1060 
1061 	beio = io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr;
1062 	softc = be_lun->softc;
1063 	memcpy(&ptrlen, io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN].bytes,
1064 	       sizeof(ptrlen));
1065 
1066 	if (ptrlen.flags != 0 || be_lun->unmap == NULL) {
1067 		ctl_free_beio(beio);
1068 		ctl_set_invalid_field(&io->scsiio,
1069 				      /*sks_valid*/ 0,
1070 				      /*command*/ 1,
1071 				      /*field*/ 0,
1072 				      /*bit_valid*/ 0,
1073 				      /*bit*/ 0);
1074 		ctl_config_write_done(io);
1075 		return;
1076 	}
1077 
1078 	/*
1079 	 * If the I/O came down with an ordered or head of queue tag, set
1080 	 * the BIO_ORDERED attribute.  For head of queue tags, that's
1081 	 * pretty much the best we can do.
1082 	 */
1083 	if ((io->scsiio.tag_type == CTL_TAG_ORDERED)
1084 	 || (io->scsiio.tag_type == CTL_TAG_HEAD_OF_QUEUE))
1085 		beio->bio_flags = BIO_ORDERED;
1086 
1087 	switch (io->scsiio.tag_type) {
1088 	case CTL_TAG_ORDERED:
1089 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1090 		break;
1091 	case CTL_TAG_HEAD_OF_QUEUE:
1092 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1093 		break;
1094 	case CTL_TAG_UNTAGGED:
1095 	case CTL_TAG_SIMPLE:
1096 	case CTL_TAG_ACA:
1097 	default:
1098 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1099 		break;
1100 	}
1101 
1102 	beio->io_len = 0;
1103 	beio->io_offset = -1;
1104 
1105 	beio->bio_cmd = BIO_DELETE;
1106 	beio->ds_trans_type = DEVSTAT_FREE;
1107 
1108 	DPRINTF("WRITE SAME at LBA %jx len %u\n",
1109 	       (uintmax_t)lbalen.lba, lbalen.len);
1110 
1111 	be_lun->unmap(be_lun, beio);
1112 }
1113 
1114 static void
1115 ctl_be_block_cw_done(struct ctl_be_block_io *beio)
1116 {
1117 	union ctl_io *io;
1118 
1119 	io = beio->io;
1120 	ctl_free_beio(beio);
1121 	ctl_config_write_done(io);
1122 }
1123 
1124 static void
1125 ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
1126 			 union ctl_io *io)
1127 {
1128 	struct ctl_be_block_io *beio;
1129 	struct ctl_be_block_softc *softc;
1130 
1131 	DPRINTF("entered\n");
1132 
1133 	softc = be_lun->softc;
1134 	beio = ctl_alloc_beio(softc);
1135 	beio->io = io;
1136 	beio->lun = be_lun;
1137 	beio->beio_cont = ctl_be_block_cw_done;
1138 	io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr = beio;
1139 
1140 	switch (io->scsiio.cdb[0]) {
1141 	case SYNCHRONIZE_CACHE:
1142 	case SYNCHRONIZE_CACHE_16:
1143 		beio->bio_cmd = BIO_FLUSH;
1144 		beio->ds_trans_type = DEVSTAT_NO_DATA;
1145 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1146 		beio->io_len = 0;
1147 		be_lun->lun_flush(be_lun, beio);
1148 		break;
1149 	case WRITE_SAME_10:
1150 	case WRITE_SAME_16:
1151 		ctl_be_block_cw_dispatch_ws(be_lun, io);
1152 		break;
1153 	case UNMAP:
1154 		ctl_be_block_cw_dispatch_unmap(be_lun, io);
1155 		break;
1156 	default:
1157 		panic("Unhandled CDB type %#x", io->scsiio.cdb[0]);
1158 		break;
1159 	}
1160 }
1161 
1162 SDT_PROBE_DEFINE1(cbb, kernel, read, start, "uint64_t");
1163 SDT_PROBE_DEFINE1(cbb, kernel, write, start, "uint64_t");
1164 SDT_PROBE_DEFINE1(cbb, kernel, read, alloc_done, "uint64_t");
1165 SDT_PROBE_DEFINE1(cbb, kernel, write, alloc_done, "uint64_t");
1166 
1167 static void
1168 ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
1169 			   union ctl_io *io)
1170 {
1171 	struct ctl_be_block_io *beio;
1172 	struct ctl_be_block_softc *softc;
1173 	struct ctl_lba_len lbalen;
1174 	uint64_t len_left, io_size_bytes;
1175 	int i;
1176 
1177 	softc = be_lun->softc;
1178 
1179 	DPRINTF("entered\n");
1180 
1181 	if ((io->io_hdr.flags & CTL_FLAG_DATA_MASK) == CTL_FLAG_DATA_IN) {
1182 		SDT_PROBE(cbb, kernel, read, start, 0, 0, 0, 0, 0);
1183 	} else {
1184 		SDT_PROBE(cbb, kernel, write, start, 0, 0, 0, 0, 0);
1185 	}
1186 
1187 	memcpy(&lbalen, io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN].bytes,
1188 	       sizeof(lbalen));
1189 
1190 	io_size_bytes = lbalen.len * be_lun->blocksize;
1191 
1192 	/*
1193 	 * XXX KDM this is temporary, until we implement chaining of beio
1194 	 * structures and multiple datamove calls to move all the data in
1195 	 * or out.
1196 	 */
1197 	if (io_size_bytes > CTLBLK_MAX_IO_SIZE) {
1198 		printf("%s: IO length %ju > max io size %u\n", __func__,
1199 		       io_size_bytes, CTLBLK_MAX_IO_SIZE);
1200 		ctl_set_invalid_field(&io->scsiio,
1201 				      /*sks_valid*/ 0,
1202 				      /*command*/ 1,
1203 				      /*field*/ 0,
1204 				      /*bit_valid*/ 0,
1205 				      /*bit*/ 0);
1206 		ctl_done(io);
1207 		return;
1208 	}
1209 
1210 	beio = ctl_alloc_beio(softc);
1211 	beio->io = io;
1212 	beio->lun = be_lun;
1213 	io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr = beio;
1214 
1215 	/*
1216 	 * If the I/O came down with an ordered or head of queue tag, set
1217 	 * the BIO_ORDERED attribute.  For head of queue tags, that's
1218 	 * pretty much the best we can do.
1219 	 *
1220 	 * XXX KDM we don't have a great way to easily know about the FUA
1221 	 * bit right now (it is decoded in ctl_read_write(), but we don't
1222 	 * pass that knowledge to the backend), and in any case we would
1223 	 * need to determine how to handle it.
1224 	 */
1225 	if ((io->scsiio.tag_type == CTL_TAG_ORDERED)
1226 	 || (io->scsiio.tag_type == CTL_TAG_HEAD_OF_QUEUE))
1227 		beio->bio_flags = BIO_ORDERED;
1228 
1229 	switch (io->scsiio.tag_type) {
1230 	case CTL_TAG_ORDERED:
1231 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1232 		break;
1233 	case CTL_TAG_HEAD_OF_QUEUE:
1234 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1235 		break;
1236 	case CTL_TAG_UNTAGGED:
1237 	case CTL_TAG_SIMPLE:
1238 	case CTL_TAG_ACA:
1239 	default:
1240 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1241 		break;
1242 	}
1243 
1244 	/*
1245 	 * This path handles read and write only.  The config write path
1246 	 * handles flush operations.
1247 	 */
1248 	if ((io->io_hdr.flags & CTL_FLAG_DATA_MASK) == CTL_FLAG_DATA_IN) {
1249 		beio->bio_cmd = BIO_READ;
1250 		beio->ds_trans_type = DEVSTAT_READ;
1251 	} else {
1252 		beio->bio_cmd = BIO_WRITE;
1253 		beio->ds_trans_type = DEVSTAT_WRITE;
1254 	}
1255 
1256 	beio->io_len = lbalen.len * be_lun->blocksize;
1257 	beio->io_offset = lbalen.lba * be_lun->blocksize;
1258 
1259 	DPRINTF("%s at LBA %jx len %u\n",
1260 	       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE",
1261 	       (uintmax_t)lbalen.lba, lbalen.len);
1262 
1263 	for (i = 0, len_left = io_size_bytes; i < CTLBLK_MAX_SEGS &&
1264 	     len_left > 0; i++) {
1265 
1266 		/*
1267 		 * Setup the S/G entry for this chunk.
1268 		 */
1269 		beio->sg_segs[i].len = min(MAXPHYS, len_left);
1270 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1271 
1272 		DPRINTF("segment %d addr %p len %zd\n", i,
1273 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1274 
1275 		beio->num_segs++;
1276 		len_left -= beio->sg_segs[i].len;
1277 	}
1278 
1279 	/*
1280 	 * For the read case, we need to read the data into our buffers and
1281 	 * then we can send it back to the user.  For the write case, we
1282 	 * need to get the data from the user first.
1283 	 */
1284 	if (beio->bio_cmd == BIO_READ) {
1285 		SDT_PROBE(cbb, kernel, read, alloc_done, 0, 0, 0, 0, 0);
1286 		be_lun->dispatch(be_lun, beio);
1287 	} else {
1288 		SDT_PROBE(cbb, kernel, write, alloc_done, 0, 0, 0, 0, 0);
1289 		io->scsiio.be_move_done = ctl_be_block_move_done;
1290 		io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
1291 		io->scsiio.kern_data_len = beio->io_len;
1292 		io->scsiio.kern_total_len = beio->io_len;
1293 		io->scsiio.kern_rel_offset = 0;
1294 		io->scsiio.kern_data_resid = 0;
1295 		io->scsiio.kern_sg_entries = beio->num_segs;
1296 		io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
1297 #ifdef CTL_TIME_IO
1298         	getbintime(&io->io_hdr.dma_start_bt);
1299 #endif
1300 		ctl_datamove(io);
1301 	}
1302 }
1303 
1304 static void
1305 ctl_be_block_worker(void *context, int pending)
1306 {
1307 	struct ctl_be_block_lun *be_lun;
1308 	struct ctl_be_block_softc *softc;
1309 	union ctl_io *io;
1310 
1311 	be_lun = (struct ctl_be_block_lun *)context;
1312 	softc = be_lun->softc;
1313 
1314 	DPRINTF("entered\n");
1315 
1316 	mtx_lock(&be_lun->lock);
1317 	for (;;) {
1318 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->datamove_queue);
1319 		if (io != NULL) {
1320 			struct ctl_be_block_io *beio;
1321 
1322 			DPRINTF("datamove queue\n");
1323 
1324 			STAILQ_REMOVE(&be_lun->datamove_queue, &io->io_hdr,
1325 				      ctl_io_hdr, links);
1326 
1327 			mtx_unlock(&be_lun->lock);
1328 
1329 			beio = (struct ctl_be_block_io *)
1330 			    io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr;
1331 
1332 			be_lun->dispatch(be_lun, beio);
1333 
1334 			mtx_lock(&be_lun->lock);
1335 			continue;
1336 		}
1337 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->config_write_queue);
1338 		if (io != NULL) {
1339 
1340 			DPRINTF("config write queue\n");
1341 
1342 			STAILQ_REMOVE(&be_lun->config_write_queue, &io->io_hdr,
1343 				      ctl_io_hdr, links);
1344 
1345 			mtx_unlock(&be_lun->lock);
1346 
1347 			ctl_be_block_cw_dispatch(be_lun, io);
1348 
1349 			mtx_lock(&be_lun->lock);
1350 			continue;
1351 		}
1352 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->input_queue);
1353 		if (io != NULL) {
1354 			DPRINTF("input queue\n");
1355 
1356 			STAILQ_REMOVE(&be_lun->input_queue, &io->io_hdr,
1357 				      ctl_io_hdr, links);
1358 			mtx_unlock(&be_lun->lock);
1359 
1360 			/*
1361 			 * We must drop the lock, since this routine and
1362 			 * its children may sleep.
1363 			 */
1364 			ctl_be_block_dispatch(be_lun, io);
1365 
1366 			mtx_lock(&be_lun->lock);
1367 			continue;
1368 		}
1369 
1370 		/*
1371 		 * If we get here, there is no work left in the queues, so
1372 		 * just break out and let the task queue go to sleep.
1373 		 */
1374 		break;
1375 	}
1376 	mtx_unlock(&be_lun->lock);
1377 }
1378 
1379 /*
1380  * Entry point from CTL to the backend for I/O.  We queue everything to a
1381  * work thread, so this just puts the I/O on a queue and wakes up the
1382  * thread.
1383  */
1384 static int
1385 ctl_be_block_submit(union ctl_io *io)
1386 {
1387 	struct ctl_be_block_lun *be_lun;
1388 	struct ctl_be_lun *ctl_be_lun;
1389 	int retval;
1390 
1391 	DPRINTF("entered\n");
1392 
1393 	retval = CTL_RETVAL_COMPLETE;
1394 
1395 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
1396 		CTL_PRIV_BACKEND_LUN].ptr;
1397 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
1398 
1399 	/*
1400 	 * Make sure we only get SCSI I/O.
1401 	 */
1402 	KASSERT(io->io_hdr.io_type == CTL_IO_SCSI, ("Non-SCSI I/O (type "
1403 		"%#x) encountered", io->io_hdr.io_type));
1404 
1405 	mtx_lock(&be_lun->lock);
1406 	/*
1407 	 * XXX KDM make sure that links is okay to use at this point.
1408 	 * Otherwise, we either need to add another field to ctl_io_hdr,
1409 	 * or deal with resource allocation here.
1410 	 */
1411 	STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1412 	mtx_unlock(&be_lun->lock);
1413 
1414 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1415 
1416 	return (retval);
1417 }
1418 
1419 static int
1420 ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
1421 			int flag, struct thread *td)
1422 {
1423 	struct ctl_be_block_softc *softc;
1424 	int error;
1425 
1426 	softc = &backend_block_softc;
1427 
1428 	error = 0;
1429 
1430 	switch (cmd) {
1431 	case CTL_LUN_REQ: {
1432 		struct ctl_lun_req *lun_req;
1433 
1434 		lun_req = (struct ctl_lun_req *)addr;
1435 
1436 		switch (lun_req->reqtype) {
1437 		case CTL_LUNREQ_CREATE:
1438 			error = ctl_be_block_create(softc, lun_req);
1439 			break;
1440 		case CTL_LUNREQ_RM:
1441 			error = ctl_be_block_rm(softc, lun_req);
1442 			break;
1443 		case CTL_LUNREQ_MODIFY:
1444 			error = ctl_be_block_modify(softc, lun_req);
1445 			break;
1446 		default:
1447 			lun_req->status = CTL_LUN_ERROR;
1448 			snprintf(lun_req->error_str, sizeof(lun_req->error_str),
1449 				 "%s: invalid LUN request type %d", __func__,
1450 				 lun_req->reqtype);
1451 			break;
1452 		}
1453 		break;
1454 	}
1455 	default:
1456 		error = ENOTTY;
1457 		break;
1458 	}
1459 
1460 	return (error);
1461 }
1462 
1463 static int
1464 ctl_be_block_open_file(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1465 {
1466 	struct ctl_be_block_filedata *file_data;
1467 	struct ctl_lun_create_params *params;
1468 	struct vattr		      vattr;
1469 	int			      error;
1470 
1471 	error = 0;
1472 	file_data = &be_lun->backend.file;
1473 	params = &req->reqdata.create;
1474 
1475 	be_lun->dev_type = CTL_BE_BLOCK_FILE;
1476 	be_lun->dispatch = ctl_be_block_dispatch_file;
1477 	be_lun->lun_flush = ctl_be_block_flush_file;
1478 
1479 	error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
1480 	if (error != 0) {
1481 		snprintf(req->error_str, sizeof(req->error_str),
1482 			 "error calling VOP_GETATTR() for file %s",
1483 			 be_lun->dev_path);
1484 		return (error);
1485 	}
1486 
1487 	/*
1488 	 * Verify that we have the ability to upgrade to exclusive
1489 	 * access on this file so we can trap errors at open instead
1490 	 * of reporting them during first access.
1491 	 */
1492 	if (VOP_ISLOCKED(be_lun->vn) != LK_EXCLUSIVE) {
1493 		vn_lock(be_lun->vn, LK_UPGRADE | LK_RETRY);
1494 		if (be_lun->vn->v_iflag & VI_DOOMED) {
1495 			error = EBADF;
1496 			snprintf(req->error_str, sizeof(req->error_str),
1497 				 "error locking file %s", be_lun->dev_path);
1498 			return (error);
1499 		}
1500 	}
1501 
1502 
1503 	file_data->cred = crhold(curthread->td_ucred);
1504 	if (params->lun_size_bytes != 0)
1505 		be_lun->size_bytes = params->lun_size_bytes;
1506 	else
1507 		be_lun->size_bytes = vattr.va_size;
1508 	/*
1509 	 * We set the multi thread flag for file operations because all
1510 	 * filesystems (in theory) are capable of allowing multiple readers
1511 	 * of a file at once.  So we want to get the maximum possible
1512 	 * concurrency.
1513 	 */
1514 	be_lun->flags |= CTL_BE_BLOCK_LUN_MULTI_THREAD;
1515 
1516 	/*
1517 	 * XXX KDM vattr.va_blocksize may be larger than 512 bytes here.
1518 	 * With ZFS, it is 131072 bytes.  Block sizes that large don't work
1519 	 * with disklabel and UFS on FreeBSD at least.  Large block sizes
1520 	 * may not work with other OSes as well.  So just export a sector
1521 	 * size of 512 bytes, which should work with any OS or
1522 	 * application.  Since our backing is a file, any block size will
1523 	 * work fine for the backing store.
1524 	 */
1525 #if 0
1526 	be_lun->blocksize= vattr.va_blocksize;
1527 #endif
1528 	if (params->blocksize_bytes != 0)
1529 		be_lun->blocksize = params->blocksize_bytes;
1530 	else
1531 		be_lun->blocksize = 512;
1532 
1533 	/*
1534 	 * Sanity check.  The media size has to be at least one
1535 	 * sector long.
1536 	 */
1537 	if (be_lun->size_bytes < be_lun->blocksize) {
1538 		error = EINVAL;
1539 		snprintf(req->error_str, sizeof(req->error_str),
1540 			 "file %s size %ju < block size %u", be_lun->dev_path,
1541 			 (uintmax_t)be_lun->size_bytes, be_lun->blocksize);
1542 	}
1543 	return (error);
1544 }
1545 
1546 static int
1547 ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1548 {
1549 	struct ctl_lun_create_params *params;
1550 	struct vattr		      vattr;
1551 	struct cdev		     *dev;
1552 	struct cdevsw		     *devsw;
1553 	int			      error;
1554 	off_t			      ps, pss, po, pos;
1555 
1556 	params = &req->reqdata.create;
1557 
1558 	be_lun->dev_type = CTL_BE_BLOCK_DEV;
1559 	be_lun->dispatch = ctl_be_block_dispatch_dev;
1560 	be_lun->lun_flush = ctl_be_block_flush_dev;
1561 	be_lun->unmap = ctl_be_block_unmap_dev;
1562 	be_lun->backend.dev.cdev = be_lun->vn->v_rdev;
1563 	be_lun->backend.dev.csw = dev_refthread(be_lun->backend.dev.cdev,
1564 					     &be_lun->backend.dev.dev_ref);
1565 	if (be_lun->backend.dev.csw == NULL)
1566 		panic("Unable to retrieve device switch");
1567 
1568 	error = VOP_GETATTR(be_lun->vn, &vattr, NOCRED);
1569 	if (error) {
1570 		snprintf(req->error_str, sizeof(req->error_str),
1571 			 "%s: error getting vnode attributes for device %s",
1572 			 __func__, be_lun->dev_path);
1573 		return (error);
1574 	}
1575 
1576 	dev = be_lun->vn->v_rdev;
1577 	devsw = dev->si_devsw;
1578 	if (!devsw->d_ioctl) {
1579 		snprintf(req->error_str, sizeof(req->error_str),
1580 			 "%s: no d_ioctl for device %s!", __func__,
1581 			 be_lun->dev_path);
1582 		return (ENODEV);
1583 	}
1584 
1585 	error = devsw->d_ioctl(dev, DIOCGSECTORSIZE,
1586 			       (caddr_t)&be_lun->blocksize, FREAD,
1587 			       curthread);
1588 	if (error) {
1589 		snprintf(req->error_str, sizeof(req->error_str),
1590 			 "%s: error %d returned for DIOCGSECTORSIZE ioctl "
1591 			 "on %s!", __func__, error, be_lun->dev_path);
1592 		return (error);
1593 	}
1594 
1595 	/*
1596 	 * If the user has asked for a blocksize that is greater than the
1597 	 * backing device's blocksize, we can do it only if the blocksize
1598 	 * the user is asking for is an even multiple of the underlying
1599 	 * device's blocksize.
1600 	 */
1601 	if ((params->blocksize_bytes != 0)
1602 	 && (params->blocksize_bytes > be_lun->blocksize)) {
1603 		uint32_t bs_multiple, tmp_blocksize;
1604 
1605 		bs_multiple = params->blocksize_bytes / be_lun->blocksize;
1606 
1607 		tmp_blocksize = bs_multiple * be_lun->blocksize;
1608 
1609 		if (tmp_blocksize == params->blocksize_bytes) {
1610 			be_lun->blocksize = params->blocksize_bytes;
1611 		} else {
1612 			snprintf(req->error_str, sizeof(req->error_str),
1613 				 "%s: requested blocksize %u is not an even "
1614 				 "multiple of backing device blocksize %u",
1615 				 __func__, params->blocksize_bytes,
1616 				 be_lun->blocksize);
1617 			return (EINVAL);
1618 
1619 		}
1620 	} else if ((params->blocksize_bytes != 0)
1621 		&& (params->blocksize_bytes != be_lun->blocksize)) {
1622 		snprintf(req->error_str, sizeof(req->error_str),
1623 			 "%s: requested blocksize %u < backing device "
1624 			 "blocksize %u", __func__, params->blocksize_bytes,
1625 			 be_lun->blocksize);
1626 		return (EINVAL);
1627 	}
1628 
1629 	error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
1630 			       (caddr_t)&be_lun->size_bytes, FREAD,
1631 			       curthread);
1632 	if (error) {
1633 		snprintf(req->error_str, sizeof(req->error_str),
1634 			 "%s: error %d returned for DIOCGMEDIASIZE "
1635 			 " ioctl on %s!", __func__, error,
1636 			 be_lun->dev_path);
1637 		return (error);
1638 	}
1639 
1640 	if (params->lun_size_bytes != 0) {
1641 		if (params->lun_size_bytes > be_lun->size_bytes) {
1642 			snprintf(req->error_str, sizeof(req->error_str),
1643 				 "%s: requested LUN size %ju > backing device "
1644 				 "size %ju", __func__,
1645 				 (uintmax_t)params->lun_size_bytes,
1646 				 (uintmax_t)be_lun->size_bytes);
1647 			return (EINVAL);
1648 		}
1649 
1650 		be_lun->size_bytes = params->lun_size_bytes;
1651 	}
1652 
1653 	error = devsw->d_ioctl(dev, DIOCGSTRIPESIZE,
1654 			       (caddr_t)&ps, FREAD, curthread);
1655 	if (error)
1656 		ps = po = 0;
1657 	else {
1658 		error = devsw->d_ioctl(dev, DIOCGSTRIPEOFFSET,
1659 				       (caddr_t)&po, FREAD, curthread);
1660 		if (error)
1661 			po = 0;
1662 	}
1663 	pss = ps / be_lun->blocksize;
1664 	pos = po / be_lun->blocksize;
1665 	if ((pss > 0) && (pss * be_lun->blocksize == ps) && (pss >= pos) &&
1666 	    ((pss & (pss - 1)) == 0) && (pos * be_lun->blocksize == po)) {
1667 		be_lun->pblockexp = fls(pss) - 1;
1668 		be_lun->pblockoff = (pss - pos) % pss;
1669 	}
1670 
1671 	return (0);
1672 }
1673 
1674 static int
1675 ctl_be_block_close(struct ctl_be_block_lun *be_lun)
1676 {
1677 	DROP_GIANT();
1678 	if (be_lun->vn) {
1679 		int flags = FREAD | FWRITE;
1680 
1681 		switch (be_lun->dev_type) {
1682 		case CTL_BE_BLOCK_DEV:
1683 			if (be_lun->backend.dev.csw) {
1684 				dev_relthread(be_lun->backend.dev.cdev,
1685 					      be_lun->backend.dev.dev_ref);
1686 				be_lun->backend.dev.csw  = NULL;
1687 				be_lun->backend.dev.cdev = NULL;
1688 			}
1689 			break;
1690 		case CTL_BE_BLOCK_FILE:
1691 			break;
1692 		case CTL_BE_BLOCK_NONE:
1693 			break;
1694 		default:
1695 			panic("Unexpected backend type.");
1696 			break;
1697 		}
1698 
1699 		(void)vn_close(be_lun->vn, flags, NOCRED, curthread);
1700 		be_lun->vn = NULL;
1701 
1702 		switch (be_lun->dev_type) {
1703 		case CTL_BE_BLOCK_DEV:
1704 			break;
1705 		case CTL_BE_BLOCK_FILE:
1706 			if (be_lun->backend.file.cred != NULL) {
1707 				crfree(be_lun->backend.file.cred);
1708 				be_lun->backend.file.cred = NULL;
1709 			}
1710 			break;
1711 		case CTL_BE_BLOCK_NONE:
1712 			break;
1713 		default:
1714 			panic("Unexpected backend type.");
1715 			break;
1716 		}
1717 	}
1718 	PICKUP_GIANT();
1719 
1720 	return (0);
1721 }
1722 
1723 static int
1724 ctl_be_block_open(struct ctl_be_block_softc *softc,
1725 		       struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1726 {
1727 	struct nameidata nd;
1728 	int		 flags;
1729 	int		 error;
1730 
1731 	/*
1732 	 * XXX KDM allow a read-only option?
1733 	 */
1734 	flags = FREAD | FWRITE;
1735 	error = 0;
1736 
1737 	if (rootvnode == NULL) {
1738 		snprintf(req->error_str, sizeof(req->error_str),
1739 			 "%s: Root filesystem is not mounted", __func__);
1740 		return (1);
1741 	}
1742 
1743 	if (!curthread->td_proc->p_fd->fd_cdir) {
1744 		curthread->td_proc->p_fd->fd_cdir = rootvnode;
1745 		VREF(rootvnode);
1746 	}
1747 	if (!curthread->td_proc->p_fd->fd_rdir) {
1748 		curthread->td_proc->p_fd->fd_rdir = rootvnode;
1749 		VREF(rootvnode);
1750 	}
1751 	if (!curthread->td_proc->p_fd->fd_jdir) {
1752 		curthread->td_proc->p_fd->fd_jdir = rootvnode;
1753 		VREF(rootvnode);
1754 	}
1755 
1756  again:
1757 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, be_lun->dev_path, curthread);
1758 	error = vn_open(&nd, &flags, 0, NULL);
1759 	if (error) {
1760 		/*
1761 		 * This is the only reasonable guess we can make as far as
1762 		 * path if the user doesn't give us a fully qualified path.
1763 		 * If they want to specify a file, they need to specify the
1764 		 * full path.
1765 		 */
1766 		if (be_lun->dev_path[0] != '/') {
1767 			char *dev_path = "/dev/";
1768 			char *dev_name;
1769 
1770 			/* Try adding device path at beginning of name */
1771 			dev_name = malloc(strlen(be_lun->dev_path)
1772 					+ strlen(dev_path) + 1,
1773 					  M_CTLBLK, M_WAITOK);
1774 			if (dev_name) {
1775 				sprintf(dev_name, "%s%s", dev_path,
1776 					be_lun->dev_path);
1777 				free(be_lun->dev_path, M_CTLBLK);
1778 				be_lun->dev_path = dev_name;
1779 				goto again;
1780 			}
1781 		}
1782 		snprintf(req->error_str, sizeof(req->error_str),
1783 			 "%s: error opening %s", __func__, be_lun->dev_path);
1784 		return (error);
1785 	}
1786 
1787 	NDFREE(&nd, NDF_ONLY_PNBUF);
1788 
1789 	be_lun->vn = nd.ni_vp;
1790 
1791 	/* We only support disks and files. */
1792 	if (vn_isdisk(be_lun->vn, &error)) {
1793 		error = ctl_be_block_open_dev(be_lun, req);
1794 	} else if (be_lun->vn->v_type == VREG) {
1795 		error = ctl_be_block_open_file(be_lun, req);
1796 	} else {
1797 		error = EINVAL;
1798 		snprintf(req->error_str, sizeof(req->error_str),
1799 			 "%s is not a disk or plain file", be_lun->dev_path);
1800 	}
1801 	VOP_UNLOCK(be_lun->vn, 0);
1802 
1803 	if (error != 0) {
1804 		ctl_be_block_close(be_lun);
1805 		return (error);
1806 	}
1807 
1808 	be_lun->blocksize_shift = fls(be_lun->blocksize) - 1;
1809 	be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
1810 
1811 	return (0);
1812 }
1813 
1814 static int
1815 ctl_be_block_create(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
1816 {
1817 	struct ctl_be_block_lun *be_lun;
1818 	struct ctl_lun_create_params *params;
1819 	struct ctl_be_arg *file_arg;
1820 	char tmpstr[32];
1821 	int retval, num_threads, unmap;
1822 	int i;
1823 
1824 	params = &req->reqdata.create;
1825 	retval = 0;
1826 
1827 	num_threads = cbb_num_threads;
1828 
1829 	file_arg = NULL;
1830 
1831 	be_lun = malloc(sizeof(*be_lun), M_CTLBLK, M_ZERO | M_WAITOK);
1832 
1833 	be_lun->softc = softc;
1834 	STAILQ_INIT(&be_lun->input_queue);
1835 	STAILQ_INIT(&be_lun->config_write_queue);
1836 	STAILQ_INIT(&be_lun->datamove_queue);
1837 	STAILQ_INIT(&be_lun->ctl_be_lun.options);
1838 	sprintf(be_lun->lunname, "cblk%d", softc->num_luns);
1839 	mtx_init(&be_lun->lock, be_lun->lunname, NULL, MTX_DEF);
1840 
1841 	be_lun->lun_zone = uma_zcreate(be_lun->lunname, MAXPHYS,
1842 	    NULL, NULL, NULL, NULL, /*align*/ 0, /*flags*/0);
1843 
1844 	if (be_lun->lun_zone == NULL) {
1845 		snprintf(req->error_str, sizeof(req->error_str),
1846 			 "%s: error allocating UMA zone", __func__);
1847 		goto bailout_error;
1848 	}
1849 
1850 	if (params->flags & CTL_LUN_FLAG_DEV_TYPE)
1851 		be_lun->ctl_be_lun.lun_type = params->device_type;
1852 	else
1853 		be_lun->ctl_be_lun.lun_type = T_DIRECT;
1854 
1855 	if (be_lun->ctl_be_lun.lun_type == T_DIRECT) {
1856 		for (i = 0; i < req->num_be_args; i++) {
1857 			if (strcmp(req->kern_be_args[i].kname, "file") == 0) {
1858 				file_arg = &req->kern_be_args[i];
1859 				break;
1860 			}
1861 		}
1862 
1863 		if (file_arg == NULL) {
1864 			snprintf(req->error_str, sizeof(req->error_str),
1865 				 "%s: no file argument specified", __func__);
1866 			goto bailout_error;
1867 		}
1868 
1869 		be_lun->dev_path = malloc(file_arg->vallen, M_CTLBLK,
1870 					  M_WAITOK | M_ZERO);
1871 
1872 		strlcpy(be_lun->dev_path, (char *)file_arg->kvalue,
1873 			file_arg->vallen);
1874 
1875 		retval = ctl_be_block_open(softc, be_lun, req);
1876 		if (retval != 0) {
1877 			retval = 0;
1878 			goto bailout_error;
1879 		}
1880 
1881 		/*
1882 		 * Tell the user the size of the file/device.
1883 		 */
1884 		params->lun_size_bytes = be_lun->size_bytes;
1885 
1886 		/*
1887 		 * The maximum LBA is the size - 1.
1888 		 */
1889 		be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
1890 	} else {
1891 		/*
1892 		 * For processor devices, we don't have any size.
1893 		 */
1894 		be_lun->blocksize = 0;
1895 		be_lun->pblockexp = 0;
1896 		be_lun->pblockoff = 0;
1897 		be_lun->size_blocks = 0;
1898 		be_lun->size_bytes = 0;
1899 		be_lun->ctl_be_lun.maxlba = 0;
1900 		params->lun_size_bytes = 0;
1901 
1902 		/*
1903 		 * Default to just 1 thread for processor devices.
1904 		 */
1905 		num_threads = 1;
1906 	}
1907 
1908 	/*
1909 	 * XXX This searching loop might be refactored to be combined with
1910 	 * the loop above,
1911 	 */
1912 	unmap = 0;
1913 	for (i = 0; i < req->num_be_args; i++) {
1914 		if (strcmp(req->kern_be_args[i].kname, "num_threads") == 0) {
1915 			struct ctl_be_arg *thread_arg;
1916 			char num_thread_str[16];
1917 			int tmp_num_threads;
1918 
1919 
1920 			thread_arg = &req->kern_be_args[i];
1921 
1922 			strlcpy(num_thread_str, (char *)thread_arg->kvalue,
1923 				min(thread_arg->vallen,
1924 				sizeof(num_thread_str)));
1925 
1926 			tmp_num_threads = strtol(num_thread_str, NULL, 0);
1927 
1928 			/*
1929 			 * We don't let the user specify less than one
1930 			 * thread, but hope he's clueful enough not to
1931 			 * specify 1000 threads.
1932 			 */
1933 			if (tmp_num_threads < 1) {
1934 				snprintf(req->error_str, sizeof(req->error_str),
1935 					 "%s: invalid number of threads %s",
1936 				         __func__, num_thread_str);
1937 				goto bailout_error;
1938 			}
1939 
1940 			num_threads = tmp_num_threads;
1941 		} else if (strcmp(req->kern_be_args[i].kname, "unmap") == 0 &&
1942 		    strcmp(req->kern_be_args[i].kvalue, "on") == 0) {
1943 			unmap = 1;
1944 		} else if (strcmp(req->kern_be_args[i].kname, "file") != 0 &&
1945 		    strcmp(req->kern_be_args[i].kname, "dev") != 0) {
1946 			struct ctl_be_lun_option *opt;
1947 
1948 			opt = malloc(sizeof(*opt), M_CTLBLK, M_WAITOK);
1949 			opt->name = malloc(strlen(req->kern_be_args[i].kname) + 1, M_CTLBLK, M_WAITOK);
1950 			strcpy(opt->name, req->kern_be_args[i].kname);
1951 			opt->value = malloc(strlen(req->kern_be_args[i].kvalue) + 1, M_CTLBLK, M_WAITOK);
1952 			strcpy(opt->value, req->kern_be_args[i].kvalue);
1953 			STAILQ_INSERT_TAIL(&be_lun->ctl_be_lun.options, opt, links);
1954 		}
1955 	}
1956 
1957 	be_lun->flags = CTL_BE_BLOCK_LUN_UNCONFIGURED;
1958 	be_lun->ctl_be_lun.flags = CTL_LUN_FLAG_PRIMARY;
1959 	if (unmap)
1960 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_UNMAP;
1961 	be_lun->ctl_be_lun.be_lun = be_lun;
1962 	be_lun->ctl_be_lun.blocksize = be_lun->blocksize;
1963 	be_lun->ctl_be_lun.pblockexp = be_lun->pblockexp;
1964 	be_lun->ctl_be_lun.pblockoff = be_lun->pblockoff;
1965 	/* Tell the user the blocksize we ended up using */
1966 	params->blocksize_bytes = be_lun->blocksize;
1967 	if (params->flags & CTL_LUN_FLAG_ID_REQ) {
1968 		be_lun->ctl_be_lun.req_lun_id = params->req_lun_id;
1969 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_ID_REQ;
1970 	} else
1971 		be_lun->ctl_be_lun.req_lun_id = 0;
1972 
1973 	be_lun->ctl_be_lun.lun_shutdown = ctl_be_block_lun_shutdown;
1974 	be_lun->ctl_be_lun.lun_config_status =
1975 		ctl_be_block_lun_config_status;
1976 	be_lun->ctl_be_lun.be = &ctl_be_block_driver;
1977 
1978 	if ((params->flags & CTL_LUN_FLAG_SERIAL_NUM) == 0) {
1979 		snprintf(tmpstr, sizeof(tmpstr), "MYSERIAL%4d",
1980 			 softc->num_luns);
1981 		strncpy((char *)be_lun->ctl_be_lun.serial_num, tmpstr,
1982 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
1983 			sizeof(tmpstr)));
1984 
1985 		/* Tell the user what we used for a serial number */
1986 		strncpy((char *)params->serial_num, tmpstr,
1987 			ctl_min(sizeof(params->serial_num), sizeof(tmpstr)));
1988 	} else {
1989 		strncpy((char *)be_lun->ctl_be_lun.serial_num,
1990 			params->serial_num,
1991 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
1992 			sizeof(params->serial_num)));
1993 	}
1994 	if ((params->flags & CTL_LUN_FLAG_DEVID) == 0) {
1995 		snprintf(tmpstr, sizeof(tmpstr), "MYDEVID%4d", softc->num_luns);
1996 		strncpy((char *)be_lun->ctl_be_lun.device_id, tmpstr,
1997 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
1998 			sizeof(tmpstr)));
1999 
2000 		/* Tell the user what we used for a device ID */
2001 		strncpy((char *)params->device_id, tmpstr,
2002 			ctl_min(sizeof(params->device_id), sizeof(tmpstr)));
2003 	} else {
2004 		strncpy((char *)be_lun->ctl_be_lun.device_id,
2005 			params->device_id,
2006 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
2007 				sizeof(params->device_id)));
2008 	}
2009 
2010 	TASK_INIT(&be_lun->io_task, /*priority*/0, ctl_be_block_worker, be_lun);
2011 
2012 	be_lun->io_taskqueue = taskqueue_create(be_lun->lunname, M_WAITOK,
2013 	    taskqueue_thread_enqueue, /*context*/&be_lun->io_taskqueue);
2014 
2015 	if (be_lun->io_taskqueue == NULL) {
2016 		snprintf(req->error_str, sizeof(req->error_str),
2017 			 "%s: Unable to create taskqueue", __func__);
2018 		goto bailout_error;
2019 	}
2020 
2021 	/*
2022 	 * Note that we start the same number of threads by default for
2023 	 * both the file case and the block device case.  For the file
2024 	 * case, we need multiple threads to allow concurrency, because the
2025 	 * vnode interface is designed to be a blocking interface.  For the
2026 	 * block device case, ZFS zvols at least will block the caller's
2027 	 * context in many instances, and so we need multiple threads to
2028 	 * overcome that problem.  Other block devices don't need as many
2029 	 * threads, but they shouldn't cause too many problems.
2030 	 *
2031 	 * If the user wants to just have a single thread for a block
2032 	 * device, he can specify that when the LUN is created, or change
2033 	 * the tunable/sysctl to alter the default number of threads.
2034 	 */
2035 	retval = taskqueue_start_threads(&be_lun->io_taskqueue,
2036 					 /*num threads*/num_threads,
2037 					 /*priority*/PWAIT,
2038 					 /*thread name*/
2039 					 "%s taskq", be_lun->lunname);
2040 
2041 	if (retval != 0)
2042 		goto bailout_error;
2043 
2044 	be_lun->num_threads = num_threads;
2045 
2046 	mtx_lock(&softc->lock);
2047 	softc->num_luns++;
2048 	STAILQ_INSERT_TAIL(&softc->lun_list, be_lun, links);
2049 
2050 	mtx_unlock(&softc->lock);
2051 
2052 	retval = ctl_add_lun(&be_lun->ctl_be_lun);
2053 	if (retval != 0) {
2054 		mtx_lock(&softc->lock);
2055 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2056 			      links);
2057 		softc->num_luns--;
2058 		mtx_unlock(&softc->lock);
2059 		snprintf(req->error_str, sizeof(req->error_str),
2060 			 "%s: ctl_add_lun() returned error %d, see dmesg for "
2061 			"details", __func__, retval);
2062 		retval = 0;
2063 		goto bailout_error;
2064 	}
2065 
2066 	mtx_lock(&softc->lock);
2067 
2068 	/*
2069 	 * Tell the config_status routine that we're waiting so it won't
2070 	 * clean up the LUN in the event of an error.
2071 	 */
2072 	be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
2073 
2074 	while (be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) {
2075 		retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
2076 		if (retval == EINTR)
2077 			break;
2078 	}
2079 	be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
2080 
2081 	if (be_lun->flags & CTL_BE_BLOCK_LUN_CONFIG_ERR) {
2082 		snprintf(req->error_str, sizeof(req->error_str),
2083 			 "%s: LUN configuration error, see dmesg for details",
2084 			 __func__);
2085 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2086 			      links);
2087 		softc->num_luns--;
2088 		mtx_unlock(&softc->lock);
2089 		goto bailout_error;
2090 	} else {
2091 		params->req_lun_id = be_lun->ctl_be_lun.lun_id;
2092 	}
2093 
2094 	mtx_unlock(&softc->lock);
2095 
2096 	be_lun->disk_stats = devstat_new_entry("cbb", params->req_lun_id,
2097 					       be_lun->blocksize,
2098 					       DEVSTAT_ALL_SUPPORTED,
2099 					       be_lun->ctl_be_lun.lun_type
2100 					       | DEVSTAT_TYPE_IF_OTHER,
2101 					       DEVSTAT_PRIORITY_OTHER);
2102 
2103 
2104 	req->status = CTL_LUN_OK;
2105 
2106 	return (retval);
2107 
2108 bailout_error:
2109 	req->status = CTL_LUN_ERROR;
2110 
2111 	ctl_be_block_close(be_lun);
2112 
2113 	free(be_lun->dev_path, M_CTLBLK);
2114 	free(be_lun, M_CTLBLK);
2115 
2116 	return (retval);
2117 }
2118 
2119 static int
2120 ctl_be_block_rm(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2121 {
2122 	struct ctl_lun_rm_params *params;
2123 	struct ctl_be_block_lun *be_lun;
2124 	int retval;
2125 
2126 	params = &req->reqdata.rm;
2127 
2128 	mtx_lock(&softc->lock);
2129 
2130 	be_lun = NULL;
2131 
2132 	STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2133 		if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2134 			break;
2135 	}
2136 	mtx_unlock(&softc->lock);
2137 
2138 	if (be_lun == NULL) {
2139 		snprintf(req->error_str, sizeof(req->error_str),
2140 			 "%s: LUN %u is not managed by the block backend",
2141 			 __func__, params->lun_id);
2142 		goto bailout_error;
2143 	}
2144 
2145 	retval = ctl_disable_lun(&be_lun->ctl_be_lun);
2146 
2147 	if (retval != 0) {
2148 		snprintf(req->error_str, sizeof(req->error_str),
2149 			 "%s: error %d returned from ctl_disable_lun() for "
2150 			 "LUN %d", __func__, retval, params->lun_id);
2151 		goto bailout_error;
2152 
2153 	}
2154 
2155 	retval = ctl_invalidate_lun(&be_lun->ctl_be_lun);
2156 	if (retval != 0) {
2157 		snprintf(req->error_str, sizeof(req->error_str),
2158 			 "%s: error %d returned from ctl_invalidate_lun() for "
2159 			 "LUN %d", __func__, retval, params->lun_id);
2160 		goto bailout_error;
2161 	}
2162 
2163 	mtx_lock(&softc->lock);
2164 
2165 	be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
2166 
2167 	while ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
2168                 retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
2169                 if (retval == EINTR)
2170                         break;
2171         }
2172 
2173 	be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
2174 
2175 	if ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
2176 		snprintf(req->error_str, sizeof(req->error_str),
2177 			 "%s: interrupted waiting for LUN to be freed",
2178 			 __func__);
2179 		mtx_unlock(&softc->lock);
2180 		goto bailout_error;
2181 	}
2182 
2183 	STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun, links);
2184 
2185 	softc->num_luns--;
2186 	mtx_unlock(&softc->lock);
2187 
2188 	taskqueue_drain(be_lun->io_taskqueue, &be_lun->io_task);
2189 
2190 	taskqueue_free(be_lun->io_taskqueue);
2191 
2192 	ctl_be_block_close(be_lun);
2193 
2194 	if (be_lun->disk_stats != NULL)
2195 		devstat_remove_entry(be_lun->disk_stats);
2196 
2197 	uma_zdestroy(be_lun->lun_zone);
2198 
2199 	free(be_lun->dev_path, M_CTLBLK);
2200 
2201 	free(be_lun, M_CTLBLK);
2202 
2203 	req->status = CTL_LUN_OK;
2204 
2205 	return (0);
2206 
2207 bailout_error:
2208 
2209 	req->status = CTL_LUN_ERROR;
2210 
2211 	return (0);
2212 }
2213 
2214 static int
2215 ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
2216 			 struct ctl_lun_req *req)
2217 {
2218 	struct vattr vattr;
2219 	int error;
2220 	struct ctl_lun_modify_params *params;
2221 
2222 	params = &req->reqdata.modify;
2223 
2224 	if (params->lun_size_bytes != 0) {
2225 		be_lun->size_bytes = params->lun_size_bytes;
2226 	} else  {
2227 		error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
2228 		if (error != 0) {
2229 			snprintf(req->error_str, sizeof(req->error_str),
2230 				 "error calling VOP_GETATTR() for file %s",
2231 				 be_lun->dev_path);
2232 			return (error);
2233 		}
2234 
2235 		be_lun->size_bytes = vattr.va_size;
2236 	}
2237 
2238 	return (0);
2239 }
2240 
2241 static int
2242 ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
2243 			struct ctl_lun_req *req)
2244 {
2245 	struct cdev *dev;
2246 	struct cdevsw *devsw;
2247 	int error;
2248 	struct ctl_lun_modify_params *params;
2249 	uint64_t size_bytes;
2250 
2251 	params = &req->reqdata.modify;
2252 
2253 	dev = be_lun->vn->v_rdev;
2254 	devsw = dev->si_devsw;
2255 	if (!devsw->d_ioctl) {
2256 		snprintf(req->error_str, sizeof(req->error_str),
2257 			 "%s: no d_ioctl for device %s!", __func__,
2258 			 be_lun->dev_path);
2259 		return (ENODEV);
2260 	}
2261 
2262 	error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
2263 			       (caddr_t)&size_bytes, FREAD,
2264 			       curthread);
2265 	if (error) {
2266 		snprintf(req->error_str, sizeof(req->error_str),
2267 			 "%s: error %d returned for DIOCGMEDIASIZE ioctl "
2268 			 "on %s!", __func__, error, be_lun->dev_path);
2269 		return (error);
2270 	}
2271 
2272 	if (params->lun_size_bytes != 0) {
2273 		if (params->lun_size_bytes > size_bytes) {
2274 			snprintf(req->error_str, sizeof(req->error_str),
2275 				 "%s: requested LUN size %ju > backing device "
2276 				 "size %ju", __func__,
2277 				 (uintmax_t)params->lun_size_bytes,
2278 				 (uintmax_t)size_bytes);
2279 			return (EINVAL);
2280 		}
2281 
2282 		be_lun->size_bytes = params->lun_size_bytes;
2283 	} else {
2284 		be_lun->size_bytes = size_bytes;
2285 	}
2286 
2287 	return (0);
2288 }
2289 
2290 static int
2291 ctl_be_block_modify(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2292 {
2293 	struct ctl_lun_modify_params *params;
2294 	struct ctl_be_block_lun *be_lun;
2295 	int error;
2296 
2297 	params = &req->reqdata.modify;
2298 
2299 	mtx_lock(&softc->lock);
2300 
2301 	be_lun = NULL;
2302 
2303 	STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2304 		if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2305 			break;
2306 	}
2307 	mtx_unlock(&softc->lock);
2308 
2309 	if (be_lun == NULL) {
2310 		snprintf(req->error_str, sizeof(req->error_str),
2311 			 "%s: LUN %u is not managed by the block backend",
2312 			 __func__, params->lun_id);
2313 		goto bailout_error;
2314 	}
2315 
2316 	if (params->lun_size_bytes != 0) {
2317 		if (params->lun_size_bytes < be_lun->blocksize) {
2318 			snprintf(req->error_str, sizeof(req->error_str),
2319 				"%s: LUN size %ju < blocksize %u", __func__,
2320 				params->lun_size_bytes, be_lun->blocksize);
2321 			goto bailout_error;
2322 		}
2323 	}
2324 
2325 	vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
2326 
2327 	if (be_lun->vn->v_type == VREG)
2328 		error = ctl_be_block_modify_file(be_lun, req);
2329 	else
2330 		error = ctl_be_block_modify_dev(be_lun, req);
2331 
2332 	VOP_UNLOCK(be_lun->vn, 0);
2333 
2334 	if (error != 0)
2335 		goto bailout_error;
2336 
2337 	be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
2338 
2339 	/*
2340 	 * The maximum LBA is the size - 1.
2341 	 *
2342 	 * XXX: Note that this field is being updated without locking,
2343 	 * 	which might cause problems on 32-bit architectures.
2344 	 */
2345 	be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
2346 	ctl_lun_capacity_changed(&be_lun->ctl_be_lun);
2347 
2348 	/* Tell the user the exact size we ended up using */
2349 	params->lun_size_bytes = be_lun->size_bytes;
2350 
2351 	req->status = CTL_LUN_OK;
2352 
2353 	return (0);
2354 
2355 bailout_error:
2356 	req->status = CTL_LUN_ERROR;
2357 
2358 	return (0);
2359 }
2360 
2361 static void
2362 ctl_be_block_lun_shutdown(void *be_lun)
2363 {
2364 	struct ctl_be_block_lun *lun;
2365 	struct ctl_be_block_softc *softc;
2366 
2367 	lun = (struct ctl_be_block_lun *)be_lun;
2368 
2369 	softc = lun->softc;
2370 
2371 	mtx_lock(&softc->lock);
2372 	lun->flags |= CTL_BE_BLOCK_LUN_UNCONFIGURED;
2373 	if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2374 		wakeup(lun);
2375 	mtx_unlock(&softc->lock);
2376 
2377 }
2378 
2379 static void
2380 ctl_be_block_lun_config_status(void *be_lun, ctl_lun_config_status status)
2381 {
2382 	struct ctl_be_block_lun *lun;
2383 	struct ctl_be_block_softc *softc;
2384 
2385 	lun = (struct ctl_be_block_lun *)be_lun;
2386 	softc = lun->softc;
2387 
2388 	if (status == CTL_LUN_CONFIG_OK) {
2389 		mtx_lock(&softc->lock);
2390 		lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2391 		if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2392 			wakeup(lun);
2393 		mtx_unlock(&softc->lock);
2394 
2395 		/*
2396 		 * We successfully added the LUN, attempt to enable it.
2397 		 */
2398 		if (ctl_enable_lun(&lun->ctl_be_lun) != 0) {
2399 			printf("%s: ctl_enable_lun() failed!\n", __func__);
2400 			if (ctl_invalidate_lun(&lun->ctl_be_lun) != 0) {
2401 				printf("%s: ctl_invalidate_lun() failed!\n",
2402 				       __func__);
2403 			}
2404 		}
2405 
2406 		return;
2407 	}
2408 
2409 
2410 	mtx_lock(&softc->lock);
2411 	lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2412 	lun->flags |= CTL_BE_BLOCK_LUN_CONFIG_ERR;
2413 	wakeup(lun);
2414 	mtx_unlock(&softc->lock);
2415 }
2416 
2417 
2418 static int
2419 ctl_be_block_config_write(union ctl_io *io)
2420 {
2421 	struct ctl_be_block_lun *be_lun;
2422 	struct ctl_be_lun *ctl_be_lun;
2423 	int retval;
2424 
2425 	retval = 0;
2426 
2427 	DPRINTF("entered\n");
2428 
2429 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
2430 		CTL_PRIV_BACKEND_LUN].ptr;
2431 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
2432 
2433 	switch (io->scsiio.cdb[0]) {
2434 	case SYNCHRONIZE_CACHE:
2435 	case SYNCHRONIZE_CACHE_16:
2436 	case WRITE_SAME_10:
2437 	case WRITE_SAME_16:
2438 	case UNMAP:
2439 		/*
2440 		 * The upper level CTL code will filter out any CDBs with
2441 		 * the immediate bit set and return the proper error.
2442 		 *
2443 		 * We don't really need to worry about what LBA range the
2444 		 * user asked to be synced out.  When they issue a sync
2445 		 * cache command, we'll sync out the whole thing.
2446 		 */
2447 		mtx_lock(&be_lun->lock);
2448 		STAILQ_INSERT_TAIL(&be_lun->config_write_queue, &io->io_hdr,
2449 				   links);
2450 		mtx_unlock(&be_lun->lock);
2451 		taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
2452 		break;
2453 	case START_STOP_UNIT: {
2454 		struct scsi_start_stop_unit *cdb;
2455 
2456 		cdb = (struct scsi_start_stop_unit *)io->scsiio.cdb;
2457 
2458 		if (cdb->how & SSS_START)
2459 			retval = ctl_start_lun(ctl_be_lun);
2460 		else {
2461 			retval = ctl_stop_lun(ctl_be_lun);
2462 			/*
2463 			 * XXX KDM Copan-specific offline behavior.
2464 			 * Figure out a reasonable way to port this?
2465 			 */
2466 #ifdef NEEDTOPORT
2467 			if ((retval == 0)
2468 			 && (cdb->byte2 & SSS_ONOFFLINE))
2469 				retval = ctl_lun_offline(ctl_be_lun);
2470 #endif
2471 		}
2472 
2473 		/*
2474 		 * In general, the above routines should not fail.  They
2475 		 * just set state for the LUN.  So we've got something
2476 		 * pretty wrong here if we can't start or stop the LUN.
2477 		 */
2478 		if (retval != 0) {
2479 			ctl_set_internal_failure(&io->scsiio,
2480 						 /*sks_valid*/ 1,
2481 						 /*retry_count*/ 0xf051);
2482 			retval = CTL_RETVAL_COMPLETE;
2483 		} else {
2484 			ctl_set_success(&io->scsiio);
2485 		}
2486 		ctl_config_write_done(io);
2487 		break;
2488 	}
2489 	default:
2490 		ctl_set_invalid_opcode(&io->scsiio);
2491 		ctl_config_write_done(io);
2492 		retval = CTL_RETVAL_COMPLETE;
2493 		break;
2494 	}
2495 
2496 	return (retval);
2497 
2498 }
2499 
2500 static int
2501 ctl_be_block_config_read(union ctl_io *io)
2502 {
2503 	return (0);
2504 }
2505 
2506 static int
2507 ctl_be_block_lun_info(void *be_lun, struct sbuf *sb)
2508 {
2509 	struct ctl_be_block_lun *lun;
2510 	int retval;
2511 
2512 	lun = (struct ctl_be_block_lun *)be_lun;
2513 	retval = 0;
2514 
2515 	retval = sbuf_printf(sb, "<num_threads>");
2516 
2517 	if (retval != 0)
2518 		goto bailout;
2519 
2520 	retval = sbuf_printf(sb, "%d", lun->num_threads);
2521 
2522 	if (retval != 0)
2523 		goto bailout;
2524 
2525 	retval = sbuf_printf(sb, "</num_threads>");
2526 
2527 	/*
2528 	 * For processor devices, we don't have a path variable.
2529 	 */
2530 	if ((retval != 0)
2531 	 || (lun->dev_path == NULL))
2532 		goto bailout;
2533 
2534 	retval = sbuf_printf(sb, "<file>");
2535 
2536 	if (retval != 0)
2537 		goto bailout;
2538 
2539 	retval = ctl_sbuf_printf_esc(sb, lun->dev_path);
2540 
2541 	if (retval != 0)
2542 		goto bailout;
2543 
2544 	retval = sbuf_printf(sb, "</file>\n");
2545 
2546 bailout:
2547 
2548 	return (retval);
2549 }
2550 
2551 int
2552 ctl_be_block_init(void)
2553 {
2554 	struct ctl_be_block_softc *softc;
2555 	int retval;
2556 
2557 	softc = &backend_block_softc;
2558 	retval = 0;
2559 
2560 	mtx_init(&softc->lock, "ctlblk", NULL, MTX_DEF);
2561 	beio_zone = uma_zcreate("beio", sizeof(struct ctl_be_block_io),
2562 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
2563 	STAILQ_INIT(&softc->disk_list);
2564 	STAILQ_INIT(&softc->lun_list);
2565 
2566 	return (retval);
2567 }
2568