xref: /freebsd/sys/cam/ctl/ctl_backend_block.c (revision 6574b8ed19b093f0af09501d2c9676c28993cb97)
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 1MB
93  * I/O.  If we get an I/O larger than that, we'll split it.
94  */
95 #define	CTLBLK_HALF_IO_SIZE	(512 * 1024)
96 #define	CTLBLK_MAX_IO_SIZE	(CTLBLK_HALF_IO_SIZE * 2)
97 #define	CTLBLK_MAX_SEG		MAXPHYS
98 #define	CTLBLK_HALF_SEGS	MAX(CTLBLK_HALF_IO_SIZE / CTLBLK_MAX_SEG, 1)
99 #define	CTLBLK_MAX_SEGS		(CTLBLK_HALF_SEGS * 2)
100 
101 #ifdef CTLBLK_DEBUG
102 #define DPRINTF(fmt, args...) \
103     printf("cbb(%s:%d): " fmt, __FUNCTION__, __LINE__, ##args)
104 #else
105 #define DPRINTF(fmt, args...) do {} while(0)
106 #endif
107 
108 #define PRIV(io)	\
109     ((struct ctl_ptr_len_flags *)&(io)->io_hdr.ctl_private[CTL_PRIV_BACKEND])
110 #define ARGS(io)	\
111     ((struct ctl_lba_len_flags *)&(io)->io_hdr.ctl_private[CTL_PRIV_LBA_LEN])
112 
113 SDT_PROVIDER_DEFINE(cbb);
114 
115 typedef enum {
116 	CTL_BE_BLOCK_LUN_UNCONFIGURED	= 0x01,
117 	CTL_BE_BLOCK_LUN_CONFIG_ERR	= 0x02,
118 	CTL_BE_BLOCK_LUN_WAITING	= 0x04,
119 	CTL_BE_BLOCK_LUN_MULTI_THREAD	= 0x08
120 } ctl_be_block_lun_flags;
121 
122 typedef enum {
123 	CTL_BE_BLOCK_NONE,
124 	CTL_BE_BLOCK_DEV,
125 	CTL_BE_BLOCK_FILE
126 } ctl_be_block_type;
127 
128 struct ctl_be_block_devdata {
129 	struct cdev *cdev;
130 	struct cdevsw *csw;
131 	int dev_ref;
132 };
133 
134 struct ctl_be_block_filedata {
135 	struct ucred *cred;
136 };
137 
138 union ctl_be_block_bedata {
139 	struct ctl_be_block_devdata dev;
140 	struct ctl_be_block_filedata file;
141 };
142 
143 struct ctl_be_block_io;
144 struct ctl_be_block_lun;
145 
146 typedef void (*cbb_dispatch_t)(struct ctl_be_block_lun *be_lun,
147 			       struct ctl_be_block_io *beio);
148 
149 /*
150  * Backend LUN structure.  There is a 1:1 mapping between a block device
151  * and a backend block LUN, and between a backend block LUN and a CTL LUN.
152  */
153 struct ctl_be_block_lun {
154 	struct ctl_block_disk *disk;
155 	char lunname[32];
156 	char *dev_path;
157 	ctl_be_block_type dev_type;
158 	struct vnode *vn;
159 	union ctl_be_block_bedata backend;
160 	cbb_dispatch_t dispatch;
161 	cbb_dispatch_t lun_flush;
162 	cbb_dispatch_t unmap;
163 	uma_zone_t lun_zone;
164 	uint64_t size_blocks;
165 	uint64_t size_bytes;
166 	uint32_t blocksize;
167 	int blocksize_shift;
168 	uint16_t pblockexp;
169 	uint16_t pblockoff;
170 	struct ctl_be_block_softc *softc;
171 	struct devstat *disk_stats;
172 	ctl_be_block_lun_flags flags;
173 	STAILQ_ENTRY(ctl_be_block_lun) links;
174 	struct ctl_be_lun ctl_be_lun;
175 	struct taskqueue *io_taskqueue;
176 	struct task io_task;
177 	int num_threads;
178 	STAILQ_HEAD(, ctl_io_hdr) input_queue;
179 	STAILQ_HEAD(, ctl_io_hdr) config_write_queue;
180 	STAILQ_HEAD(, ctl_io_hdr) datamove_queue;
181 	struct mtx_padalign io_lock;
182 	struct mtx_padalign queue_lock;
183 };
184 
185 /*
186  * Overall softc structure for the block backend module.
187  */
188 struct ctl_be_block_softc {
189 	struct mtx			 lock;
190 	int				 num_disks;
191 	STAILQ_HEAD(, ctl_block_disk)	 disk_list;
192 	int				 num_luns;
193 	STAILQ_HEAD(, ctl_be_block_lun)	 lun_list;
194 };
195 
196 static struct ctl_be_block_softc backend_block_softc;
197 
198 /*
199  * Per-I/O information.
200  */
201 struct ctl_be_block_io {
202 	union ctl_io			*io;
203 	struct ctl_sg_entry		sg_segs[CTLBLK_MAX_SEGS];
204 	struct iovec			xiovecs[CTLBLK_MAX_SEGS];
205 	int				bio_cmd;
206 	int				num_segs;
207 	int				num_bios_sent;
208 	int				num_bios_done;
209 	int				send_complete;
210 	int				num_errors;
211 	struct bintime			ds_t0;
212 	devstat_tag_type		ds_tag_type;
213 	devstat_trans_flags		ds_trans_type;
214 	uint64_t			io_len;
215 	uint64_t			io_offset;
216 	struct ctl_be_block_softc	*softc;
217 	struct ctl_be_block_lun		*lun;
218 	void (*beio_cont)(struct ctl_be_block_io *beio); /* to continue processing */
219 };
220 
221 static int cbb_num_threads = 14;
222 SYSCTL_NODE(_kern_cam_ctl, OID_AUTO, block, CTLFLAG_RD, 0,
223 	    "CAM Target Layer Block Backend");
224 SYSCTL_INT(_kern_cam_ctl_block, OID_AUTO, num_threads, CTLFLAG_RWTUN,
225            &cbb_num_threads, 0, "Number of threads per backing file");
226 
227 static struct ctl_be_block_io *ctl_alloc_beio(struct ctl_be_block_softc *softc);
228 static void ctl_free_beio(struct ctl_be_block_io *beio);
229 static void ctl_complete_beio(struct ctl_be_block_io *beio);
230 static int ctl_be_block_move_done(union ctl_io *io);
231 static void ctl_be_block_biodone(struct bio *bio);
232 static void ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
233 				    struct ctl_be_block_io *beio);
234 static void ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
235 				       struct ctl_be_block_io *beio);
236 static void ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
237 				   struct ctl_be_block_io *beio);
238 static void ctl_be_block_unmap_dev(struct ctl_be_block_lun *be_lun,
239 				   struct ctl_be_block_io *beio);
240 static void ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
241 				      struct ctl_be_block_io *beio);
242 static void ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
243 				    union ctl_io *io);
244 static void ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
245 				  union ctl_io *io);
246 static void ctl_be_block_worker(void *context, int pending);
247 static int ctl_be_block_submit(union ctl_io *io);
248 static int ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
249 				   int flag, struct thread *td);
250 static int ctl_be_block_open_file(struct ctl_be_block_lun *be_lun,
251 				  struct ctl_lun_req *req);
252 static int ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun,
253 				 struct ctl_lun_req *req);
254 static int ctl_be_block_close(struct ctl_be_block_lun *be_lun);
255 static int ctl_be_block_open(struct ctl_be_block_softc *softc,
256 			     struct ctl_be_block_lun *be_lun,
257 			     struct ctl_lun_req *req);
258 static int ctl_be_block_create(struct ctl_be_block_softc *softc,
259 			       struct ctl_lun_req *req);
260 static int ctl_be_block_rm(struct ctl_be_block_softc *softc,
261 			   struct ctl_lun_req *req);
262 static int ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
263 				  struct ctl_lun_req *req);
264 static int ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
265 				 struct ctl_lun_req *req);
266 static int ctl_be_block_modify(struct ctl_be_block_softc *softc,
267 			   struct ctl_lun_req *req);
268 static void ctl_be_block_lun_shutdown(void *be_lun);
269 static void ctl_be_block_lun_config_status(void *be_lun,
270 					   ctl_lun_config_status status);
271 static int ctl_be_block_config_write(union ctl_io *io);
272 static int ctl_be_block_config_read(union ctl_io *io);
273 static int ctl_be_block_lun_info(void *be_lun, struct sbuf *sb);
274 int ctl_be_block_init(void);
275 
276 static struct ctl_backend_driver ctl_be_block_driver =
277 {
278 	.name = "block",
279 	.flags = CTL_BE_FLAG_HAS_CONFIG,
280 	.init = ctl_be_block_init,
281 	.data_submit = ctl_be_block_submit,
282 	.data_move_done = ctl_be_block_move_done,
283 	.config_read = ctl_be_block_config_read,
284 	.config_write = ctl_be_block_config_write,
285 	.ioctl = ctl_be_block_ioctl,
286 	.lun_info = ctl_be_block_lun_info
287 };
288 
289 MALLOC_DEFINE(M_CTLBLK, "ctlblk", "Memory used for CTL block backend");
290 CTL_BACKEND_DECLARE(cbb, ctl_be_block_driver);
291 
292 static uma_zone_t beio_zone;
293 
294 static struct ctl_be_block_io *
295 ctl_alloc_beio(struct ctl_be_block_softc *softc)
296 {
297 	struct ctl_be_block_io *beio;
298 
299 	beio = uma_zalloc(beio_zone, M_WAITOK | M_ZERO);
300 	beio->softc = softc;
301 	return (beio);
302 }
303 
304 static void
305 ctl_free_beio(struct ctl_be_block_io *beio)
306 {
307 	int duplicate_free;
308 	int i;
309 
310 	duplicate_free = 0;
311 
312 	for (i = 0; i < beio->num_segs; i++) {
313 		if (beio->sg_segs[i].addr == NULL)
314 			duplicate_free++;
315 
316 		uma_zfree(beio->lun->lun_zone, beio->sg_segs[i].addr);
317 		beio->sg_segs[i].addr = NULL;
318 
319 		/* For compare we had two equal S/G lists. */
320 		if (ARGS(beio->io)->flags & CTL_LLF_COMPARE) {
321 			uma_zfree(beio->lun->lun_zone,
322 			    beio->sg_segs[i + CTLBLK_HALF_SEGS].addr);
323 			beio->sg_segs[i + CTLBLK_HALF_SEGS].addr = NULL;
324 		}
325 	}
326 
327 	if (duplicate_free > 0) {
328 		printf("%s: %d duplicate frees out of %d segments\n", __func__,
329 		       duplicate_free, beio->num_segs);
330 	}
331 
332 	uma_zfree(beio_zone, beio);
333 }
334 
335 static void
336 ctl_complete_beio(struct ctl_be_block_io *beio)
337 {
338 	union ctl_io *io = beio->io;
339 
340 	if (beio->beio_cont != NULL) {
341 		beio->beio_cont(beio);
342 	} else {
343 		ctl_free_beio(beio);
344 		ctl_data_submit_done(io);
345 	}
346 }
347 
348 static int
349 ctl_be_block_move_done(union ctl_io *io)
350 {
351 	struct ctl_be_block_io *beio;
352 	struct ctl_be_block_lun *be_lun;
353 	struct ctl_lba_len_flags *lbalen;
354 #ifdef CTL_TIME_IO
355 	struct bintime cur_bt;
356 #endif
357 	int i;
358 
359 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
360 	be_lun = beio->lun;
361 
362 	DPRINTF("entered\n");
363 
364 #ifdef CTL_TIME_IO
365 	getbintime(&cur_bt);
366 	bintime_sub(&cur_bt, &io->io_hdr.dma_start_bt);
367 	bintime_add(&io->io_hdr.dma_bt, &cur_bt);
368 	io->io_hdr.num_dmas++;
369 #endif
370 	io->scsiio.kern_rel_offset += io->scsiio.kern_data_len;
371 
372 	/*
373 	 * We set status at this point for read commands, and write
374 	 * commands with errors.
375 	 */
376 	if ((io->io_hdr.port_status == 0) &&
377 	    ((io->io_hdr.flags & CTL_FLAG_ABORT) == 0) &&
378 	    ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_STATUS_NONE)) {
379 		lbalen = ARGS(beio->io);
380 		if (lbalen->flags & CTL_LLF_READ) {
381 			ctl_set_success(&io->scsiio);
382 		} else if (lbalen->flags & CTL_LLF_COMPARE) {
383 			/* We have two data blocks ready for comparison. */
384 			for (i = 0; i < beio->num_segs; i++) {
385 				if (memcmp(beio->sg_segs[i].addr,
386 				    beio->sg_segs[i + CTLBLK_HALF_SEGS].addr,
387 				    beio->sg_segs[i].len) != 0)
388 					break;
389 			}
390 			if (i < beio->num_segs)
391 				ctl_set_sense(&io->scsiio,
392 				    /*current_error*/ 1,
393 				    /*sense_key*/ SSD_KEY_MISCOMPARE,
394 				    /*asc*/ 0x1D,
395 				    /*ascq*/ 0x00,
396 				    SSD_ELEM_NONE);
397 			else
398 				ctl_set_success(&io->scsiio);
399 		}
400 	}
401 	else if ((io->io_hdr.port_status != 0)
402 	      && ((io->io_hdr.flags & CTL_FLAG_ABORT) == 0)
403 	      && ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_STATUS_NONE)) {
404 		/*
405 		 * For hardware error sense keys, the sense key
406 		 * specific value is defined to be a retry count,
407 		 * but we use it to pass back an internal FETD
408 		 * error code.  XXX KDM  Hopefully the FETD is only
409 		 * using 16 bits for an error code, since that's
410 		 * all the space we have in the sks field.
411 		 */
412 		ctl_set_internal_failure(&io->scsiio,
413 					 /*sks_valid*/ 1,
414 					 /*retry_count*/
415 					 io->io_hdr.port_status);
416 	}
417 
418 	/*
419 	 * If this is a read, or a write with errors, it is done.
420 	 */
421 	if ((beio->bio_cmd == BIO_READ)
422 	 || ((io->io_hdr.flags & CTL_FLAG_ABORT) != 0)
423 	 || ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE)) {
424 		ctl_complete_beio(beio);
425 		return (0);
426 	}
427 
428 	/*
429 	 * At this point, we have a write and the DMA completed
430 	 * successfully.  We now have to queue it to the task queue to
431 	 * execute the backend I/O.  That is because we do blocking
432 	 * memory allocations, and in the file backing case, blocking I/O.
433 	 * This move done routine is generally called in the SIM's
434 	 * interrupt context, and therefore we cannot block.
435 	 */
436 	mtx_lock(&be_lun->queue_lock);
437 	/*
438 	 * XXX KDM make sure that links is okay to use at this point.
439 	 * Otherwise, we either need to add another field to ctl_io_hdr,
440 	 * or deal with resource allocation here.
441 	 */
442 	STAILQ_INSERT_TAIL(&be_lun->datamove_queue, &io->io_hdr, links);
443 	mtx_unlock(&be_lun->queue_lock);
444 
445 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
446 
447 	return (0);
448 }
449 
450 static void
451 ctl_be_block_biodone(struct bio *bio)
452 {
453 	struct ctl_be_block_io *beio;
454 	struct ctl_be_block_lun *be_lun;
455 	union ctl_io *io;
456 	int error;
457 
458 	beio = bio->bio_caller1;
459 	be_lun = beio->lun;
460 	io = beio->io;
461 
462 	DPRINTF("entered\n");
463 
464 	error = bio->bio_error;
465 	mtx_lock(&be_lun->io_lock);
466 	if (error != 0)
467 		beio->num_errors++;
468 
469 	beio->num_bios_done++;
470 
471 	/*
472 	 * XXX KDM will this cause WITNESS to complain?  Holding a lock
473 	 * during the free might cause it to complain.
474 	 */
475 	g_destroy_bio(bio);
476 
477 	/*
478 	 * If the send complete bit isn't set, or we aren't the last I/O to
479 	 * complete, then we're done.
480 	 */
481 	if ((beio->send_complete == 0)
482 	 || (beio->num_bios_done < beio->num_bios_sent)) {
483 		mtx_unlock(&be_lun->io_lock);
484 		return;
485 	}
486 
487 	/*
488 	 * At this point, we've verified that we are the last I/O to
489 	 * complete, so it's safe to drop the lock.
490 	 */
491 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
492 	    beio->ds_tag_type, beio->ds_trans_type,
493 	    /*now*/ NULL, /*then*/&beio->ds_t0);
494 	mtx_unlock(&be_lun->io_lock);
495 
496 	/*
497 	 * If there are any errors from the backing device, we fail the
498 	 * entire I/O with a medium error.
499 	 */
500 	if (beio->num_errors > 0) {
501 		if (error == EOPNOTSUPP) {
502 			ctl_set_invalid_opcode(&io->scsiio);
503 		} else if (beio->bio_cmd == BIO_FLUSH) {
504 			/* XXX KDM is there is a better error here? */
505 			ctl_set_internal_failure(&io->scsiio,
506 						 /*sks_valid*/ 1,
507 						 /*retry_count*/ 0xbad2);
508 		} else
509 			ctl_set_medium_error(&io->scsiio);
510 		ctl_complete_beio(beio);
511 		return;
512 	}
513 
514 	/*
515 	 * If this is a write, a flush, a delete or verify, we're all done.
516 	 * If this is a read, we can now send the data to the user.
517 	 */
518 	if ((beio->bio_cmd == BIO_WRITE)
519 	 || (beio->bio_cmd == BIO_FLUSH)
520 	 || (beio->bio_cmd == BIO_DELETE)
521 	 || (ARGS(io)->flags & CTL_LLF_VERIFY)) {
522 		ctl_set_success(&io->scsiio);
523 		ctl_complete_beio(beio);
524 	} else {
525 #ifdef CTL_TIME_IO
526         	getbintime(&io->io_hdr.dma_start_bt);
527 #endif
528 		ctl_datamove(io);
529 	}
530 }
531 
532 static void
533 ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
534 			struct ctl_be_block_io *beio)
535 {
536 	union ctl_io *io = beio->io;
537 	struct mount *mountpoint;
538 	int error, lock_flags;
539 
540 	DPRINTF("entered\n");
541 
542 	binuptime(&beio->ds_t0);
543 	mtx_lock(&be_lun->io_lock);
544 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
545 	mtx_unlock(&be_lun->io_lock);
546 
547 	(void) vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
548 
549 	if (MNT_SHARED_WRITES(mountpoint)
550 	 || ((mountpoint == NULL)
551 	  && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
552 		lock_flags = LK_SHARED;
553 	else
554 		lock_flags = LK_EXCLUSIVE;
555 
556 	vn_lock(be_lun->vn, lock_flags | LK_RETRY);
557 
558 	error = VOP_FSYNC(be_lun->vn, MNT_WAIT, curthread);
559 	VOP_UNLOCK(be_lun->vn, 0);
560 
561 	vn_finished_write(mountpoint);
562 
563 	mtx_lock(&be_lun->io_lock);
564 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
565 	    beio->ds_tag_type, beio->ds_trans_type,
566 	    /*now*/ NULL, /*then*/&beio->ds_t0);
567 	mtx_unlock(&be_lun->io_lock);
568 
569 	if (error == 0)
570 		ctl_set_success(&io->scsiio);
571 	else {
572 		/* XXX KDM is there is a better error here? */
573 		ctl_set_internal_failure(&io->scsiio,
574 					 /*sks_valid*/ 1,
575 					 /*retry_count*/ 0xbad1);
576 	}
577 
578 	ctl_complete_beio(beio);
579 }
580 
581 SDT_PROBE_DEFINE1(cbb, kernel, read, file_start, "uint64_t");
582 SDT_PROBE_DEFINE1(cbb, kernel, write, file_start, "uint64_t");
583 SDT_PROBE_DEFINE1(cbb, kernel, read, file_done,"uint64_t");
584 SDT_PROBE_DEFINE1(cbb, kernel, write, file_done, "uint64_t");
585 
586 static void
587 ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
588 			   struct ctl_be_block_io *beio)
589 {
590 	struct ctl_be_block_filedata *file_data;
591 	union ctl_io *io;
592 	struct uio xuio;
593 	struct iovec *xiovec;
594 	int flags;
595 	int error, i;
596 
597 	DPRINTF("entered\n");
598 
599 	file_data = &be_lun->backend.file;
600 	io = beio->io;
601 	flags = 0;
602 	if (ARGS(io)->flags & CTL_LLF_DPO)
603 		flags |= IO_DIRECT;
604 	if (beio->bio_cmd == BIO_WRITE && ARGS(io)->flags & CTL_LLF_FUA)
605 		flags |= IO_SYNC;
606 
607 	bzero(&xuio, sizeof(xuio));
608 	if (beio->bio_cmd == BIO_READ) {
609 		SDT_PROBE(cbb, kernel, read, file_start, 0, 0, 0, 0, 0);
610 		xuio.uio_rw = UIO_READ;
611 	} else {
612 		SDT_PROBE(cbb, kernel, write, file_start, 0, 0, 0, 0, 0);
613 		xuio.uio_rw = UIO_WRITE;
614 	}
615 	xuio.uio_offset = beio->io_offset;
616 	xuio.uio_resid = beio->io_len;
617 	xuio.uio_segflg = UIO_SYSSPACE;
618 	xuio.uio_iov = beio->xiovecs;
619 	xuio.uio_iovcnt = beio->num_segs;
620 	xuio.uio_td = curthread;
621 
622 	for (i = 0, xiovec = xuio.uio_iov; i < xuio.uio_iovcnt; i++, xiovec++) {
623 		xiovec->iov_base = beio->sg_segs[i].addr;
624 		xiovec->iov_len = beio->sg_segs[i].len;
625 	}
626 
627 	binuptime(&beio->ds_t0);
628 	mtx_lock(&be_lun->io_lock);
629 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
630 	mtx_unlock(&be_lun->io_lock);
631 
632 	if (beio->bio_cmd == BIO_READ) {
633 		vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
634 
635 		/*
636 		 * UFS pays attention to IO_DIRECT for reads.  If the
637 		 * DIRECTIO option is configured into the kernel, it calls
638 		 * ffs_rawread().  But that only works for single-segment
639 		 * uios with user space addresses.  In our case, with a
640 		 * kernel uio, it still reads into the buffer cache, but it
641 		 * will just try to release the buffer from the cache later
642 		 * on in ffs_read().
643 		 *
644 		 * ZFS does not pay attention to IO_DIRECT for reads.
645 		 *
646 		 * UFS does not pay attention to IO_SYNC for reads.
647 		 *
648 		 * ZFS pays attention to IO_SYNC (which translates into the
649 		 * Solaris define FRSYNC for zfs_read()) for reads.  It
650 		 * attempts to sync the file before reading.
651 		 *
652 		 * So, to attempt to provide some barrier semantics in the
653 		 * BIO_ORDERED case, set both IO_DIRECT and IO_SYNC.
654 		 */
655 		error = VOP_READ(be_lun->vn, &xuio, flags, file_data->cred);
656 
657 		VOP_UNLOCK(be_lun->vn, 0);
658 		SDT_PROBE(cbb, kernel, read, file_done, 0, 0, 0, 0, 0);
659 	} else {
660 		struct mount *mountpoint;
661 		int lock_flags;
662 
663 		(void)vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
664 
665 		if (MNT_SHARED_WRITES(mountpoint)
666 		 || ((mountpoint == NULL)
667 		  && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
668 			lock_flags = LK_SHARED;
669 		else
670 			lock_flags = LK_EXCLUSIVE;
671 
672 		vn_lock(be_lun->vn, lock_flags | LK_RETRY);
673 
674 		/*
675 		 * UFS pays attention to IO_DIRECT for writes.  The write
676 		 * is done asynchronously.  (Normally the write would just
677 		 * get put into cache.
678 		 *
679 		 * UFS pays attention to IO_SYNC for writes.  It will
680 		 * attempt to write the buffer out synchronously if that
681 		 * flag is set.
682 		 *
683 		 * ZFS does not pay attention to IO_DIRECT for writes.
684 		 *
685 		 * ZFS pays attention to IO_SYNC (a.k.a. FSYNC or FRSYNC)
686 		 * for writes.  It will flush the transaction from the
687 		 * cache before returning.
688 		 *
689 		 * So if we've got the BIO_ORDERED flag set, we want
690 		 * IO_SYNC in either the UFS or ZFS case.
691 		 */
692 		error = VOP_WRITE(be_lun->vn, &xuio, flags, file_data->cred);
693 		VOP_UNLOCK(be_lun->vn, 0);
694 
695 		vn_finished_write(mountpoint);
696 		SDT_PROBE(cbb, kernel, write, file_done, 0, 0, 0, 0, 0);
697         }
698 
699 	mtx_lock(&be_lun->io_lock);
700 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
701 	    beio->ds_tag_type, beio->ds_trans_type,
702 	    /*now*/ NULL, /*then*/&beio->ds_t0);
703 	mtx_unlock(&be_lun->io_lock);
704 
705 	/*
706 	 * If we got an error, set the sense data to "MEDIUM ERROR" and
707 	 * return the I/O to the user.
708 	 */
709 	if (error != 0) {
710 		char path_str[32];
711 
712 		ctl_scsi_path_string(io, path_str, sizeof(path_str));
713 		/*
714 		 * XXX KDM ZFS returns ENOSPC when the underlying
715 		 * filesystem fills up.  What kind of SCSI error should we
716 		 * return for that?
717 		 */
718 		printf("%s%s command returned errno %d\n", path_str,
719 		       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE", error);
720 		ctl_set_medium_error(&io->scsiio);
721 		ctl_complete_beio(beio);
722 		return;
723 	}
724 
725 	/*
726 	 * If this is a write or a verify, we're all done.
727 	 * If this is a read, we can now send the data to the user.
728 	 */
729 	if ((beio->bio_cmd == BIO_WRITE) ||
730 	    (ARGS(io)->flags & CTL_LLF_VERIFY)) {
731 		ctl_set_success(&io->scsiio);
732 		ctl_complete_beio(beio);
733 	} else {
734 #ifdef CTL_TIME_IO
735         	getbintime(&io->io_hdr.dma_start_bt);
736 #endif
737 		ctl_datamove(io);
738 	}
739 }
740 
741 static void
742 ctl_be_block_dispatch_zvol(struct ctl_be_block_lun *be_lun,
743 			   struct ctl_be_block_io *beio)
744 {
745 	struct ctl_be_block_devdata *dev_data;
746 	union ctl_io *io;
747 	struct uio xuio;
748 	struct iovec *xiovec;
749 	int flags;
750 	int error, i;
751 
752 	DPRINTF("entered\n");
753 
754 	dev_data = &be_lun->backend.dev;
755 	io = beio->io;
756 	flags = 0;
757 	if (ARGS(io)->flags & CTL_LLF_DPO)
758 		flags |= IO_DIRECT;
759 	if (beio->bio_cmd == BIO_WRITE && ARGS(io)->flags & CTL_LLF_FUA)
760 		flags |= IO_SYNC;
761 
762 	bzero(&xuio, sizeof(xuio));
763 	if (beio->bio_cmd == BIO_READ) {
764 		SDT_PROBE(cbb, kernel, read, file_start, 0, 0, 0, 0, 0);
765 		xuio.uio_rw = UIO_READ;
766 	} else {
767 		SDT_PROBE(cbb, kernel, write, file_start, 0, 0, 0, 0, 0);
768 		xuio.uio_rw = UIO_WRITE;
769 	}
770 	xuio.uio_offset = beio->io_offset;
771 	xuio.uio_resid = beio->io_len;
772 	xuio.uio_segflg = UIO_SYSSPACE;
773 	xuio.uio_iov = beio->xiovecs;
774 	xuio.uio_iovcnt = beio->num_segs;
775 	xuio.uio_td = curthread;
776 
777 	for (i = 0, xiovec = xuio.uio_iov; i < xuio.uio_iovcnt; i++, xiovec++) {
778 		xiovec->iov_base = beio->sg_segs[i].addr;
779 		xiovec->iov_len = beio->sg_segs[i].len;
780 	}
781 
782 	binuptime(&beio->ds_t0);
783 	mtx_lock(&be_lun->io_lock);
784 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
785 	mtx_unlock(&be_lun->io_lock);
786 
787 	if (beio->bio_cmd == BIO_READ) {
788 		error = (*dev_data->csw->d_read)(dev_data->cdev, &xuio, flags);
789 		SDT_PROBE(cbb, kernel, read, file_done, 0, 0, 0, 0, 0);
790 	} else {
791 		error = (*dev_data->csw->d_write)(dev_data->cdev, &xuio, flags);
792 		SDT_PROBE(cbb, kernel, write, file_done, 0, 0, 0, 0, 0);
793 	}
794 
795 	mtx_lock(&be_lun->io_lock);
796 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
797 	    beio->ds_tag_type, beio->ds_trans_type,
798 	    /*now*/ NULL, /*then*/&beio->ds_t0);
799 	mtx_unlock(&be_lun->io_lock);
800 
801 	/*
802 	 * If we got an error, set the sense data to "MEDIUM ERROR" and
803 	 * return the I/O to the user.
804 	 */
805 	if (error != 0) {
806 		ctl_set_medium_error(&io->scsiio);
807 		ctl_complete_beio(beio);
808 		return;
809 	}
810 
811 	/*
812 	 * If this is a write or a verify, we're all done.
813 	 * If this is a read, we can now send the data to the user.
814 	 */
815 	if ((beio->bio_cmd == BIO_WRITE) ||
816 	    (ARGS(io)->flags & CTL_LLF_VERIFY)) {
817 		ctl_set_success(&io->scsiio);
818 		ctl_complete_beio(beio);
819 	} else {
820 #ifdef CTL_TIME_IO
821         	getbintime(&io->io_hdr.dma_start_bt);
822 #endif
823 		ctl_datamove(io);
824 	}
825 }
826 
827 static void
828 ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
829 		       struct ctl_be_block_io *beio)
830 {
831 	struct bio *bio;
832 	union ctl_io *io;
833 	struct ctl_be_block_devdata *dev_data;
834 
835 	dev_data = &be_lun->backend.dev;
836 	io = beio->io;
837 
838 	DPRINTF("entered\n");
839 
840 	/* This can't fail, it's a blocking allocation. */
841 	bio = g_alloc_bio();
842 
843 	bio->bio_cmd	    = BIO_FLUSH;
844 	bio->bio_flags	   |= BIO_ORDERED;
845 	bio->bio_dev	    = dev_data->cdev;
846 	bio->bio_offset	    = 0;
847 	bio->bio_data	    = 0;
848 	bio->bio_done	    = ctl_be_block_biodone;
849 	bio->bio_caller1    = beio;
850 	bio->bio_pblkno	    = 0;
851 
852 	/*
853 	 * We don't need to acquire the LUN lock here, because we are only
854 	 * sending one bio, and so there is no other context to synchronize
855 	 * with.
856 	 */
857 	beio->num_bios_sent = 1;
858 	beio->send_complete = 1;
859 
860 	binuptime(&beio->ds_t0);
861 	mtx_lock(&be_lun->io_lock);
862 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
863 	mtx_unlock(&be_lun->io_lock);
864 
865 	(*dev_data->csw->d_strategy)(bio);
866 }
867 
868 static void
869 ctl_be_block_unmap_dev_range(struct ctl_be_block_lun *be_lun,
870 		       struct ctl_be_block_io *beio,
871 		       uint64_t off, uint64_t len, int last)
872 {
873 	struct bio *bio;
874 	struct ctl_be_block_devdata *dev_data;
875 	uint64_t maxlen;
876 
877 	dev_data = &be_lun->backend.dev;
878 	maxlen = LONG_MAX - (LONG_MAX % be_lun->blocksize);
879 	while (len > 0) {
880 		bio = g_alloc_bio();
881 		bio->bio_cmd	    = BIO_DELETE;
882 		bio->bio_dev	    = dev_data->cdev;
883 		bio->bio_offset	    = off;
884 		bio->bio_length	    = MIN(len, maxlen);
885 		bio->bio_data	    = 0;
886 		bio->bio_done	    = ctl_be_block_biodone;
887 		bio->bio_caller1    = beio;
888 		bio->bio_pblkno     = off / be_lun->blocksize;
889 
890 		off += bio->bio_length;
891 		len -= bio->bio_length;
892 
893 		mtx_lock(&be_lun->io_lock);
894 		beio->num_bios_sent++;
895 		if (last && len == 0)
896 			beio->send_complete = 1;
897 		mtx_unlock(&be_lun->io_lock);
898 
899 		(*dev_data->csw->d_strategy)(bio);
900 	}
901 }
902 
903 static void
904 ctl_be_block_unmap_dev(struct ctl_be_block_lun *be_lun,
905 		       struct ctl_be_block_io *beio)
906 {
907 	union ctl_io *io;
908 	struct ctl_be_block_devdata *dev_data;
909 	struct ctl_ptr_len_flags *ptrlen;
910 	struct scsi_unmap_desc *buf, *end;
911 	uint64_t len;
912 
913 	dev_data = &be_lun->backend.dev;
914 	io = beio->io;
915 
916 	DPRINTF("entered\n");
917 
918 	binuptime(&beio->ds_t0);
919 	mtx_lock(&be_lun->io_lock);
920 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
921 	mtx_unlock(&be_lun->io_lock);
922 
923 	if (beio->io_offset == -1) {
924 		beio->io_len = 0;
925 		ptrlen = (struct ctl_ptr_len_flags *)&io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN];
926 		buf = (struct scsi_unmap_desc *)ptrlen->ptr;
927 		end = buf + ptrlen->len / sizeof(*buf);
928 		for (; buf < end; buf++) {
929 			len = (uint64_t)scsi_4btoul(buf->length) *
930 			    be_lun->blocksize;
931 			beio->io_len += len;
932 			ctl_be_block_unmap_dev_range(be_lun, beio,
933 			    scsi_8btou64(buf->lba) * be_lun->blocksize, len,
934 			    (end - buf < 2) ? TRUE : FALSE);
935 		}
936 	} else
937 		ctl_be_block_unmap_dev_range(be_lun, beio,
938 		    beio->io_offset, beio->io_len, TRUE);
939 }
940 
941 static void
942 ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
943 			  struct ctl_be_block_io *beio)
944 {
945 	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
946 	int i;
947 	struct bio *bio;
948 	struct ctl_be_block_devdata *dev_data;
949 	off_t cur_offset;
950 	int max_iosize;
951 
952 	DPRINTF("entered\n");
953 
954 	dev_data = &be_lun->backend.dev;
955 
956 	/*
957 	 * We have to limit our I/O size to the maximum supported by the
958 	 * backend device.  Hopefully it is MAXPHYS.  If the driver doesn't
959 	 * set it properly, use DFLTPHYS.
960 	 */
961 	max_iosize = dev_data->cdev->si_iosize_max;
962 	if (max_iosize < PAGE_SIZE)
963 		max_iosize = DFLTPHYS;
964 
965 	cur_offset = beio->io_offset;
966 	for (i = 0; i < beio->num_segs; i++) {
967 		size_t cur_size;
968 		uint8_t *cur_ptr;
969 
970 		cur_size = beio->sg_segs[i].len;
971 		cur_ptr = beio->sg_segs[i].addr;
972 
973 		while (cur_size > 0) {
974 			/* This can't fail, it's a blocking allocation. */
975 			bio = g_alloc_bio();
976 
977 			KASSERT(bio != NULL, ("g_alloc_bio() failed!\n"));
978 
979 			bio->bio_cmd = beio->bio_cmd;
980 			bio->bio_dev = dev_data->cdev;
981 			bio->bio_caller1 = beio;
982 			bio->bio_length = min(cur_size, max_iosize);
983 			bio->bio_offset = cur_offset;
984 			bio->bio_data = cur_ptr;
985 			bio->bio_done = ctl_be_block_biodone;
986 			bio->bio_pblkno = cur_offset / be_lun->blocksize;
987 
988 			cur_offset += bio->bio_length;
989 			cur_ptr += bio->bio_length;
990 			cur_size -= bio->bio_length;
991 
992 			TAILQ_INSERT_TAIL(&queue, bio, bio_queue);
993 			beio->num_bios_sent++;
994 		}
995 	}
996 	binuptime(&beio->ds_t0);
997 	mtx_lock(&be_lun->io_lock);
998 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
999 	beio->send_complete = 1;
1000 	mtx_unlock(&be_lun->io_lock);
1001 
1002 	/*
1003 	 * Fire off all allocated requests!
1004 	 */
1005 	while ((bio = TAILQ_FIRST(&queue)) != NULL) {
1006 		TAILQ_REMOVE(&queue, bio, bio_queue);
1007 		(*dev_data->csw->d_strategy)(bio);
1008 	}
1009 }
1010 
1011 static void
1012 ctl_be_block_cw_done_ws(struct ctl_be_block_io *beio)
1013 {
1014 	union ctl_io *io;
1015 
1016 	io = beio->io;
1017 	ctl_free_beio(beio);
1018 	if ((io->io_hdr.flags & CTL_FLAG_ABORT) ||
1019 	    ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE &&
1020 	     (io->io_hdr.status & CTL_STATUS_MASK) != CTL_SUCCESS)) {
1021 		ctl_config_write_done(io);
1022 		return;
1023 	}
1024 
1025 	ctl_be_block_config_write(io);
1026 }
1027 
1028 static void
1029 ctl_be_block_cw_dispatch_ws(struct ctl_be_block_lun *be_lun,
1030 			    union ctl_io *io)
1031 {
1032 	struct ctl_be_block_io *beio;
1033 	struct ctl_be_block_softc *softc;
1034 	struct ctl_lba_len_flags *lbalen;
1035 	uint64_t len_left, lba;
1036 	int i, seglen;
1037 	uint8_t *buf, *end;
1038 
1039 	DPRINTF("entered\n");
1040 
1041 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1042 	softc = be_lun->softc;
1043 	lbalen = ARGS(beio->io);
1044 
1045 	if (lbalen->flags & ~(SWS_LBDATA | SWS_UNMAP | SWS_ANCHOR) ||
1046 	    (lbalen->flags & (SWS_UNMAP | SWS_ANCHOR) && be_lun->unmap == NULL)) {
1047 		ctl_free_beio(beio);
1048 		ctl_set_invalid_field(&io->scsiio,
1049 				      /*sks_valid*/ 1,
1050 				      /*command*/ 1,
1051 				      /*field*/ 1,
1052 				      /*bit_valid*/ 0,
1053 				      /*bit*/ 0);
1054 		ctl_config_write_done(io);
1055 		return;
1056 	}
1057 
1058 	switch (io->scsiio.tag_type) {
1059 	case CTL_TAG_ORDERED:
1060 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1061 		break;
1062 	case CTL_TAG_HEAD_OF_QUEUE:
1063 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1064 		break;
1065 	case CTL_TAG_UNTAGGED:
1066 	case CTL_TAG_SIMPLE:
1067 	case CTL_TAG_ACA:
1068 	default:
1069 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1070 		break;
1071 	}
1072 
1073 	if (lbalen->flags & (SWS_UNMAP | SWS_ANCHOR)) {
1074 		beio->io_offset = lbalen->lba * be_lun->blocksize;
1075 		beio->io_len = (uint64_t)lbalen->len * be_lun->blocksize;
1076 		beio->bio_cmd = BIO_DELETE;
1077 		beio->ds_trans_type = DEVSTAT_FREE;
1078 
1079 		be_lun->unmap(be_lun, beio);
1080 		return;
1081 	}
1082 
1083 	beio->bio_cmd = BIO_WRITE;
1084 	beio->ds_trans_type = DEVSTAT_WRITE;
1085 
1086 	DPRINTF("WRITE SAME at LBA %jx len %u\n",
1087 	       (uintmax_t)lbalen->lba, lbalen->len);
1088 
1089 	len_left = (uint64_t)lbalen->len * be_lun->blocksize;
1090 	for (i = 0, lba = 0; i < CTLBLK_MAX_SEGS && len_left > 0; i++) {
1091 
1092 		/*
1093 		 * Setup the S/G entry for this chunk.
1094 		 */
1095 		seglen = MIN(CTLBLK_MAX_SEG, len_left);
1096 		seglen -= seglen % be_lun->blocksize;
1097 		beio->sg_segs[i].len = seglen;
1098 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1099 
1100 		DPRINTF("segment %d addr %p len %zd\n", i,
1101 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1102 
1103 		beio->num_segs++;
1104 		len_left -= seglen;
1105 
1106 		buf = beio->sg_segs[i].addr;
1107 		end = buf + seglen;
1108 		for (; buf < end; buf += be_lun->blocksize) {
1109 			memcpy(buf, io->scsiio.kern_data_ptr, be_lun->blocksize);
1110 			if (lbalen->flags & SWS_LBDATA)
1111 				scsi_ulto4b(lbalen->lba + lba, buf);
1112 			lba++;
1113 		}
1114 	}
1115 
1116 	beio->io_offset = lbalen->lba * be_lun->blocksize;
1117 	beio->io_len = lba * be_lun->blocksize;
1118 
1119 	/* We can not do all in one run. Correct and schedule rerun. */
1120 	if (len_left > 0) {
1121 		lbalen->lba += lba;
1122 		lbalen->len -= lba;
1123 		beio->beio_cont = ctl_be_block_cw_done_ws;
1124 	}
1125 
1126 	be_lun->dispatch(be_lun, beio);
1127 }
1128 
1129 static void
1130 ctl_be_block_cw_dispatch_unmap(struct ctl_be_block_lun *be_lun,
1131 			    union ctl_io *io)
1132 {
1133 	struct ctl_be_block_io *beio;
1134 	struct ctl_be_block_softc *softc;
1135 	struct ctl_ptr_len_flags *ptrlen;
1136 
1137 	DPRINTF("entered\n");
1138 
1139 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1140 	softc = be_lun->softc;
1141 	ptrlen = (struct ctl_ptr_len_flags *)&io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN];
1142 
1143 	if ((ptrlen->flags & ~SU_ANCHOR) != 0 || be_lun->unmap == NULL) {
1144 		ctl_free_beio(beio);
1145 		ctl_set_invalid_field(&io->scsiio,
1146 				      /*sks_valid*/ 0,
1147 				      /*command*/ 1,
1148 				      /*field*/ 0,
1149 				      /*bit_valid*/ 0,
1150 				      /*bit*/ 0);
1151 		ctl_config_write_done(io);
1152 		return;
1153 	}
1154 
1155 	switch (io->scsiio.tag_type) {
1156 	case CTL_TAG_ORDERED:
1157 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1158 		break;
1159 	case CTL_TAG_HEAD_OF_QUEUE:
1160 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1161 		break;
1162 	case CTL_TAG_UNTAGGED:
1163 	case CTL_TAG_SIMPLE:
1164 	case CTL_TAG_ACA:
1165 	default:
1166 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1167 		break;
1168 	}
1169 
1170 	beio->io_len = 0;
1171 	beio->io_offset = -1;
1172 
1173 	beio->bio_cmd = BIO_DELETE;
1174 	beio->ds_trans_type = DEVSTAT_FREE;
1175 
1176 	DPRINTF("UNMAP\n");
1177 
1178 	be_lun->unmap(be_lun, beio);
1179 }
1180 
1181 static void
1182 ctl_be_block_cw_done(struct ctl_be_block_io *beio)
1183 {
1184 	union ctl_io *io;
1185 
1186 	io = beio->io;
1187 	ctl_free_beio(beio);
1188 	ctl_config_write_done(io);
1189 }
1190 
1191 static void
1192 ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
1193 			 union ctl_io *io)
1194 {
1195 	struct ctl_be_block_io *beio;
1196 	struct ctl_be_block_softc *softc;
1197 
1198 	DPRINTF("entered\n");
1199 
1200 	softc = be_lun->softc;
1201 	beio = ctl_alloc_beio(softc);
1202 	beio->io = io;
1203 	beio->lun = be_lun;
1204 	beio->beio_cont = ctl_be_block_cw_done;
1205 	PRIV(io)->ptr = (void *)beio;
1206 
1207 	switch (io->scsiio.cdb[0]) {
1208 	case SYNCHRONIZE_CACHE:
1209 	case SYNCHRONIZE_CACHE_16:
1210 		beio->bio_cmd = BIO_FLUSH;
1211 		beio->ds_trans_type = DEVSTAT_NO_DATA;
1212 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1213 		beio->io_len = 0;
1214 		be_lun->lun_flush(be_lun, beio);
1215 		break;
1216 	case WRITE_SAME_10:
1217 	case WRITE_SAME_16:
1218 		ctl_be_block_cw_dispatch_ws(be_lun, io);
1219 		break;
1220 	case UNMAP:
1221 		ctl_be_block_cw_dispatch_unmap(be_lun, io);
1222 		break;
1223 	default:
1224 		panic("Unhandled CDB type %#x", io->scsiio.cdb[0]);
1225 		break;
1226 	}
1227 }
1228 
1229 SDT_PROBE_DEFINE1(cbb, kernel, read, start, "uint64_t");
1230 SDT_PROBE_DEFINE1(cbb, kernel, write, start, "uint64_t");
1231 SDT_PROBE_DEFINE1(cbb, kernel, read, alloc_done, "uint64_t");
1232 SDT_PROBE_DEFINE1(cbb, kernel, write, alloc_done, "uint64_t");
1233 
1234 static void
1235 ctl_be_block_next(struct ctl_be_block_io *beio)
1236 {
1237 	struct ctl_be_block_lun *be_lun;
1238 	union ctl_io *io;
1239 
1240 	io = beio->io;
1241 	be_lun = beio->lun;
1242 	ctl_free_beio(beio);
1243 	if ((io->io_hdr.flags & CTL_FLAG_ABORT) ||
1244 	    ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE &&
1245 	     (io->io_hdr.status & CTL_STATUS_MASK) != CTL_SUCCESS)) {
1246 		ctl_data_submit_done(io);
1247 		return;
1248 	}
1249 
1250 	io->io_hdr.status &= ~CTL_STATUS_MASK;
1251 	io->io_hdr.status |= CTL_STATUS_NONE;
1252 
1253 	mtx_lock(&be_lun->queue_lock);
1254 	/*
1255 	 * XXX KDM make sure that links is okay to use at this point.
1256 	 * Otherwise, we either need to add another field to ctl_io_hdr,
1257 	 * or deal with resource allocation here.
1258 	 */
1259 	STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1260 	mtx_unlock(&be_lun->queue_lock);
1261 
1262 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1263 }
1264 
1265 static void
1266 ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
1267 			   union ctl_io *io)
1268 {
1269 	struct ctl_be_block_io *beio;
1270 	struct ctl_be_block_softc *softc;
1271 	struct ctl_lba_len_flags *lbalen;
1272 	struct ctl_ptr_len_flags *bptrlen;
1273 	uint64_t len_left, lbas;
1274 	int i;
1275 
1276 	softc = be_lun->softc;
1277 
1278 	DPRINTF("entered\n");
1279 
1280 	lbalen = ARGS(io);
1281 	if (lbalen->flags & CTL_LLF_WRITE) {
1282 		SDT_PROBE(cbb, kernel, write, start, 0, 0, 0, 0, 0);
1283 	} else {
1284 		SDT_PROBE(cbb, kernel, read, start, 0, 0, 0, 0, 0);
1285 	}
1286 
1287 	beio = ctl_alloc_beio(softc);
1288 	beio->io = io;
1289 	beio->lun = be_lun;
1290 	bptrlen = PRIV(io);
1291 	bptrlen->ptr = (void *)beio;
1292 
1293 	switch (io->scsiio.tag_type) {
1294 	case CTL_TAG_ORDERED:
1295 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1296 		break;
1297 	case CTL_TAG_HEAD_OF_QUEUE:
1298 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1299 		break;
1300 	case CTL_TAG_UNTAGGED:
1301 	case CTL_TAG_SIMPLE:
1302 	case CTL_TAG_ACA:
1303 	default:
1304 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1305 		break;
1306 	}
1307 
1308 	if (lbalen->flags & CTL_LLF_WRITE) {
1309 		beio->bio_cmd = BIO_WRITE;
1310 		beio->ds_trans_type = DEVSTAT_WRITE;
1311 	} else {
1312 		beio->bio_cmd = BIO_READ;
1313 		beio->ds_trans_type = DEVSTAT_READ;
1314 	}
1315 
1316 	DPRINTF("%s at LBA %jx len %u @%ju\n",
1317 	       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE",
1318 	       (uintmax_t)lbalen->lba, lbalen->len, bptrlen->len);
1319 	if (lbalen->flags & CTL_LLF_COMPARE)
1320 		lbas = CTLBLK_HALF_IO_SIZE;
1321 	else
1322 		lbas = CTLBLK_MAX_IO_SIZE;
1323 	lbas = MIN(lbalen->len - bptrlen->len, lbas / be_lun->blocksize);
1324 	beio->io_offset = (lbalen->lba + bptrlen->len) * be_lun->blocksize;
1325 	beio->io_len = lbas * be_lun->blocksize;
1326 	bptrlen->len += lbas;
1327 
1328 	for (i = 0, len_left = beio->io_len; len_left > 0; i++) {
1329 		KASSERT(i < CTLBLK_MAX_SEGS, ("Too many segs (%d >= %d)",
1330 		    i, CTLBLK_MAX_SEGS));
1331 
1332 		/*
1333 		 * Setup the S/G entry for this chunk.
1334 		 */
1335 		beio->sg_segs[i].len = min(CTLBLK_MAX_SEG, len_left);
1336 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1337 
1338 		DPRINTF("segment %d addr %p len %zd\n", i,
1339 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1340 
1341 		/* Set up second segment for compare operation. */
1342 		if (lbalen->flags & CTL_LLF_COMPARE) {
1343 			beio->sg_segs[i + CTLBLK_HALF_SEGS].len =
1344 			    beio->sg_segs[i].len;
1345 			beio->sg_segs[i + CTLBLK_HALF_SEGS].addr =
1346 			    uma_zalloc(be_lun->lun_zone, M_WAITOK);
1347 		}
1348 
1349 		beio->num_segs++;
1350 		len_left -= beio->sg_segs[i].len;
1351 	}
1352 	if (bptrlen->len < lbalen->len)
1353 		beio->beio_cont = ctl_be_block_next;
1354 	io->scsiio.be_move_done = ctl_be_block_move_done;
1355 	/* For compare we have separate S/G lists for read and datamove. */
1356 	if (lbalen->flags & CTL_LLF_COMPARE)
1357 		io->scsiio.kern_data_ptr = (uint8_t *)&beio->sg_segs[CTLBLK_HALF_SEGS];
1358 	else
1359 		io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
1360 	io->scsiio.kern_data_len = beio->io_len;
1361 	io->scsiio.kern_data_resid = 0;
1362 	io->scsiio.kern_sg_entries = beio->num_segs;
1363 	io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
1364 
1365 	/*
1366 	 * For the read case, we need to read the data into our buffers and
1367 	 * then we can send it back to the user.  For the write case, we
1368 	 * need to get the data from the user first.
1369 	 */
1370 	if (beio->bio_cmd == BIO_READ) {
1371 		SDT_PROBE(cbb, kernel, read, alloc_done, 0, 0, 0, 0, 0);
1372 		be_lun->dispatch(be_lun, beio);
1373 	} else {
1374 		SDT_PROBE(cbb, kernel, write, alloc_done, 0, 0, 0, 0, 0);
1375 #ifdef CTL_TIME_IO
1376         	getbintime(&io->io_hdr.dma_start_bt);
1377 #endif
1378 		ctl_datamove(io);
1379 	}
1380 }
1381 
1382 static void
1383 ctl_be_block_worker(void *context, int pending)
1384 {
1385 	struct ctl_be_block_lun *be_lun;
1386 	struct ctl_be_block_softc *softc;
1387 	union ctl_io *io;
1388 
1389 	be_lun = (struct ctl_be_block_lun *)context;
1390 	softc = be_lun->softc;
1391 
1392 	DPRINTF("entered\n");
1393 
1394 	mtx_lock(&be_lun->queue_lock);
1395 	for (;;) {
1396 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->datamove_queue);
1397 		if (io != NULL) {
1398 			struct ctl_be_block_io *beio;
1399 
1400 			DPRINTF("datamove queue\n");
1401 
1402 			STAILQ_REMOVE(&be_lun->datamove_queue, &io->io_hdr,
1403 				      ctl_io_hdr, links);
1404 
1405 			mtx_unlock(&be_lun->queue_lock);
1406 
1407 			beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1408 
1409 			be_lun->dispatch(be_lun, beio);
1410 
1411 			mtx_lock(&be_lun->queue_lock);
1412 			continue;
1413 		}
1414 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->config_write_queue);
1415 		if (io != NULL) {
1416 
1417 			DPRINTF("config write queue\n");
1418 
1419 			STAILQ_REMOVE(&be_lun->config_write_queue, &io->io_hdr,
1420 				      ctl_io_hdr, links);
1421 
1422 			mtx_unlock(&be_lun->queue_lock);
1423 
1424 			ctl_be_block_cw_dispatch(be_lun, io);
1425 
1426 			mtx_lock(&be_lun->queue_lock);
1427 			continue;
1428 		}
1429 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->input_queue);
1430 		if (io != NULL) {
1431 			DPRINTF("input queue\n");
1432 
1433 			STAILQ_REMOVE(&be_lun->input_queue, &io->io_hdr,
1434 				      ctl_io_hdr, links);
1435 			mtx_unlock(&be_lun->queue_lock);
1436 
1437 			/*
1438 			 * We must drop the lock, since this routine and
1439 			 * its children may sleep.
1440 			 */
1441 			ctl_be_block_dispatch(be_lun, io);
1442 
1443 			mtx_lock(&be_lun->queue_lock);
1444 			continue;
1445 		}
1446 
1447 		/*
1448 		 * If we get here, there is no work left in the queues, so
1449 		 * just break out and let the task queue go to sleep.
1450 		 */
1451 		break;
1452 	}
1453 	mtx_unlock(&be_lun->queue_lock);
1454 }
1455 
1456 /*
1457  * Entry point from CTL to the backend for I/O.  We queue everything to a
1458  * work thread, so this just puts the I/O on a queue and wakes up the
1459  * thread.
1460  */
1461 static int
1462 ctl_be_block_submit(union ctl_io *io)
1463 {
1464 	struct ctl_be_block_lun *be_lun;
1465 	struct ctl_be_lun *ctl_be_lun;
1466 
1467 	DPRINTF("entered\n");
1468 
1469 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
1470 		CTL_PRIV_BACKEND_LUN].ptr;
1471 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
1472 
1473 	/*
1474 	 * Make sure we only get SCSI I/O.
1475 	 */
1476 	KASSERT(io->io_hdr.io_type == CTL_IO_SCSI, ("Non-SCSI I/O (type "
1477 		"%#x) encountered", io->io_hdr.io_type));
1478 
1479 	PRIV(io)->len = 0;
1480 
1481 	mtx_lock(&be_lun->queue_lock);
1482 	/*
1483 	 * XXX KDM make sure that links is okay to use at this point.
1484 	 * Otherwise, we either need to add another field to ctl_io_hdr,
1485 	 * or deal with resource allocation here.
1486 	 */
1487 	STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1488 	mtx_unlock(&be_lun->queue_lock);
1489 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1490 
1491 	return (CTL_RETVAL_COMPLETE);
1492 }
1493 
1494 static int
1495 ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
1496 			int flag, struct thread *td)
1497 {
1498 	struct ctl_be_block_softc *softc;
1499 	int error;
1500 
1501 	softc = &backend_block_softc;
1502 
1503 	error = 0;
1504 
1505 	switch (cmd) {
1506 	case CTL_LUN_REQ: {
1507 		struct ctl_lun_req *lun_req;
1508 
1509 		lun_req = (struct ctl_lun_req *)addr;
1510 
1511 		switch (lun_req->reqtype) {
1512 		case CTL_LUNREQ_CREATE:
1513 			error = ctl_be_block_create(softc, lun_req);
1514 			break;
1515 		case CTL_LUNREQ_RM:
1516 			error = ctl_be_block_rm(softc, lun_req);
1517 			break;
1518 		case CTL_LUNREQ_MODIFY:
1519 			error = ctl_be_block_modify(softc, lun_req);
1520 			break;
1521 		default:
1522 			lun_req->status = CTL_LUN_ERROR;
1523 			snprintf(lun_req->error_str, sizeof(lun_req->error_str),
1524 				 "%s: invalid LUN request type %d", __func__,
1525 				 lun_req->reqtype);
1526 			break;
1527 		}
1528 		break;
1529 	}
1530 	default:
1531 		error = ENOTTY;
1532 		break;
1533 	}
1534 
1535 	return (error);
1536 }
1537 
1538 static int
1539 ctl_be_block_open_file(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1540 {
1541 	struct ctl_be_block_filedata *file_data;
1542 	struct ctl_lun_create_params *params;
1543 	struct vattr		      vattr;
1544 	int			      error;
1545 
1546 	error = 0;
1547 	file_data = &be_lun->backend.file;
1548 	params = &req->reqdata.create;
1549 
1550 	be_lun->dev_type = CTL_BE_BLOCK_FILE;
1551 	be_lun->dispatch = ctl_be_block_dispatch_file;
1552 	be_lun->lun_flush = ctl_be_block_flush_file;
1553 
1554 	error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
1555 	if (error != 0) {
1556 		snprintf(req->error_str, sizeof(req->error_str),
1557 			 "error calling VOP_GETATTR() for file %s",
1558 			 be_lun->dev_path);
1559 		return (error);
1560 	}
1561 
1562 	/*
1563 	 * Verify that we have the ability to upgrade to exclusive
1564 	 * access on this file so we can trap errors at open instead
1565 	 * of reporting them during first access.
1566 	 */
1567 	if (VOP_ISLOCKED(be_lun->vn) != LK_EXCLUSIVE) {
1568 		vn_lock(be_lun->vn, LK_UPGRADE | LK_RETRY);
1569 		if (be_lun->vn->v_iflag & VI_DOOMED) {
1570 			error = EBADF;
1571 			snprintf(req->error_str, sizeof(req->error_str),
1572 				 "error locking file %s", be_lun->dev_path);
1573 			return (error);
1574 		}
1575 	}
1576 
1577 
1578 	file_data->cred = crhold(curthread->td_ucred);
1579 	if (params->lun_size_bytes != 0)
1580 		be_lun->size_bytes = params->lun_size_bytes;
1581 	else
1582 		be_lun->size_bytes = vattr.va_size;
1583 	/*
1584 	 * We set the multi thread flag for file operations because all
1585 	 * filesystems (in theory) are capable of allowing multiple readers
1586 	 * of a file at once.  So we want to get the maximum possible
1587 	 * concurrency.
1588 	 */
1589 	be_lun->flags |= CTL_BE_BLOCK_LUN_MULTI_THREAD;
1590 
1591 	/*
1592 	 * XXX KDM vattr.va_blocksize may be larger than 512 bytes here.
1593 	 * With ZFS, it is 131072 bytes.  Block sizes that large don't work
1594 	 * with disklabel and UFS on FreeBSD at least.  Large block sizes
1595 	 * may not work with other OSes as well.  So just export a sector
1596 	 * size of 512 bytes, which should work with any OS or
1597 	 * application.  Since our backing is a file, any block size will
1598 	 * work fine for the backing store.
1599 	 */
1600 #if 0
1601 	be_lun->blocksize= vattr.va_blocksize;
1602 #endif
1603 	if (params->blocksize_bytes != 0)
1604 		be_lun->blocksize = params->blocksize_bytes;
1605 	else
1606 		be_lun->blocksize = 512;
1607 
1608 	/*
1609 	 * Sanity check.  The media size has to be at least one
1610 	 * sector long.
1611 	 */
1612 	if (be_lun->size_bytes < be_lun->blocksize) {
1613 		error = EINVAL;
1614 		snprintf(req->error_str, sizeof(req->error_str),
1615 			 "file %s size %ju < block size %u", be_lun->dev_path,
1616 			 (uintmax_t)be_lun->size_bytes, be_lun->blocksize);
1617 	}
1618 	return (error);
1619 }
1620 
1621 static int
1622 ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1623 {
1624 	struct ctl_lun_create_params *params;
1625 	struct vattr		      vattr;
1626 	struct cdev		     *dev;
1627 	struct cdevsw		     *devsw;
1628 	int			      error;
1629 	off_t			      ps, pss, po, pos;
1630 
1631 	params = &req->reqdata.create;
1632 
1633 	be_lun->dev_type = CTL_BE_BLOCK_DEV;
1634 	be_lun->backend.dev.cdev = be_lun->vn->v_rdev;
1635 	be_lun->backend.dev.csw = dev_refthread(be_lun->backend.dev.cdev,
1636 					     &be_lun->backend.dev.dev_ref);
1637 	if (be_lun->backend.dev.csw == NULL)
1638 		panic("Unable to retrieve device switch");
1639 	if (strcmp(be_lun->backend.dev.csw->d_name, "zvol") == 0)
1640 		be_lun->dispatch = ctl_be_block_dispatch_zvol;
1641 	else
1642 		be_lun->dispatch = ctl_be_block_dispatch_dev;
1643 	be_lun->lun_flush = ctl_be_block_flush_dev;
1644 	be_lun->unmap = ctl_be_block_unmap_dev;
1645 
1646 	error = VOP_GETATTR(be_lun->vn, &vattr, NOCRED);
1647 	if (error) {
1648 		snprintf(req->error_str, sizeof(req->error_str),
1649 			 "%s: error getting vnode attributes for device %s",
1650 			 __func__, be_lun->dev_path);
1651 		return (error);
1652 	}
1653 
1654 	dev = be_lun->vn->v_rdev;
1655 	devsw = dev->si_devsw;
1656 	if (!devsw->d_ioctl) {
1657 		snprintf(req->error_str, sizeof(req->error_str),
1658 			 "%s: no d_ioctl for device %s!", __func__,
1659 			 be_lun->dev_path);
1660 		return (ENODEV);
1661 	}
1662 
1663 	error = devsw->d_ioctl(dev, DIOCGSECTORSIZE,
1664 			       (caddr_t)&be_lun->blocksize, FREAD,
1665 			       curthread);
1666 	if (error) {
1667 		snprintf(req->error_str, sizeof(req->error_str),
1668 			 "%s: error %d returned for DIOCGSECTORSIZE ioctl "
1669 			 "on %s!", __func__, error, be_lun->dev_path);
1670 		return (error);
1671 	}
1672 
1673 	/*
1674 	 * If the user has asked for a blocksize that is greater than the
1675 	 * backing device's blocksize, we can do it only if the blocksize
1676 	 * the user is asking for is an even multiple of the underlying
1677 	 * device's blocksize.
1678 	 */
1679 	if ((params->blocksize_bytes != 0)
1680 	 && (params->blocksize_bytes > be_lun->blocksize)) {
1681 		uint32_t bs_multiple, tmp_blocksize;
1682 
1683 		bs_multiple = params->blocksize_bytes / be_lun->blocksize;
1684 
1685 		tmp_blocksize = bs_multiple * be_lun->blocksize;
1686 
1687 		if (tmp_blocksize == params->blocksize_bytes) {
1688 			be_lun->blocksize = params->blocksize_bytes;
1689 		} else {
1690 			snprintf(req->error_str, sizeof(req->error_str),
1691 				 "%s: requested blocksize %u is not an even "
1692 				 "multiple of backing device blocksize %u",
1693 				 __func__, params->blocksize_bytes,
1694 				 be_lun->blocksize);
1695 			return (EINVAL);
1696 
1697 		}
1698 	} else if ((params->blocksize_bytes != 0)
1699 		&& (params->blocksize_bytes != be_lun->blocksize)) {
1700 		snprintf(req->error_str, sizeof(req->error_str),
1701 			 "%s: requested blocksize %u < backing device "
1702 			 "blocksize %u", __func__, params->blocksize_bytes,
1703 			 be_lun->blocksize);
1704 		return (EINVAL);
1705 	}
1706 
1707 	error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
1708 			       (caddr_t)&be_lun->size_bytes, FREAD,
1709 			       curthread);
1710 	if (error) {
1711 		snprintf(req->error_str, sizeof(req->error_str),
1712 			 "%s: error %d returned for DIOCGMEDIASIZE "
1713 			 " ioctl on %s!", __func__, error,
1714 			 be_lun->dev_path);
1715 		return (error);
1716 	}
1717 
1718 	if (params->lun_size_bytes != 0) {
1719 		if (params->lun_size_bytes > be_lun->size_bytes) {
1720 			snprintf(req->error_str, sizeof(req->error_str),
1721 				 "%s: requested LUN size %ju > backing device "
1722 				 "size %ju", __func__,
1723 				 (uintmax_t)params->lun_size_bytes,
1724 				 (uintmax_t)be_lun->size_bytes);
1725 			return (EINVAL);
1726 		}
1727 
1728 		be_lun->size_bytes = params->lun_size_bytes;
1729 	}
1730 
1731 	error = devsw->d_ioctl(dev, DIOCGSTRIPESIZE,
1732 			       (caddr_t)&ps, FREAD, curthread);
1733 	if (error)
1734 		ps = po = 0;
1735 	else {
1736 		error = devsw->d_ioctl(dev, DIOCGSTRIPEOFFSET,
1737 				       (caddr_t)&po, FREAD, curthread);
1738 		if (error)
1739 			po = 0;
1740 	}
1741 	pss = ps / be_lun->blocksize;
1742 	pos = po / be_lun->blocksize;
1743 	if ((pss > 0) && (pss * be_lun->blocksize == ps) && (pss >= pos) &&
1744 	    ((pss & (pss - 1)) == 0) && (pos * be_lun->blocksize == po)) {
1745 		be_lun->pblockexp = fls(pss) - 1;
1746 		be_lun->pblockoff = (pss - pos) % pss;
1747 	}
1748 
1749 	return (0);
1750 }
1751 
1752 static int
1753 ctl_be_block_close(struct ctl_be_block_lun *be_lun)
1754 {
1755 	DROP_GIANT();
1756 	if (be_lun->vn) {
1757 		int flags = FREAD | FWRITE;
1758 
1759 		switch (be_lun->dev_type) {
1760 		case CTL_BE_BLOCK_DEV:
1761 			if (be_lun->backend.dev.csw) {
1762 				dev_relthread(be_lun->backend.dev.cdev,
1763 					      be_lun->backend.dev.dev_ref);
1764 				be_lun->backend.dev.csw  = NULL;
1765 				be_lun->backend.dev.cdev = NULL;
1766 			}
1767 			break;
1768 		case CTL_BE_BLOCK_FILE:
1769 			break;
1770 		case CTL_BE_BLOCK_NONE:
1771 			break;
1772 		default:
1773 			panic("Unexpected backend type.");
1774 			break;
1775 		}
1776 
1777 		(void)vn_close(be_lun->vn, flags, NOCRED, curthread);
1778 		be_lun->vn = NULL;
1779 
1780 		switch (be_lun->dev_type) {
1781 		case CTL_BE_BLOCK_DEV:
1782 			break;
1783 		case CTL_BE_BLOCK_FILE:
1784 			if (be_lun->backend.file.cred != NULL) {
1785 				crfree(be_lun->backend.file.cred);
1786 				be_lun->backend.file.cred = NULL;
1787 			}
1788 			break;
1789 		case CTL_BE_BLOCK_NONE:
1790 			break;
1791 		default:
1792 			panic("Unexpected backend type.");
1793 			break;
1794 		}
1795 	}
1796 	PICKUP_GIANT();
1797 
1798 	return (0);
1799 }
1800 
1801 static int
1802 ctl_be_block_open(struct ctl_be_block_softc *softc,
1803 		       struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1804 {
1805 	struct nameidata nd;
1806 	int		 flags;
1807 	int		 error;
1808 
1809 	/*
1810 	 * XXX KDM allow a read-only option?
1811 	 */
1812 	flags = FREAD | FWRITE;
1813 	error = 0;
1814 
1815 	if (rootvnode == NULL) {
1816 		snprintf(req->error_str, sizeof(req->error_str),
1817 			 "%s: Root filesystem is not mounted", __func__);
1818 		return (1);
1819 	}
1820 
1821 	if (!curthread->td_proc->p_fd->fd_cdir) {
1822 		curthread->td_proc->p_fd->fd_cdir = rootvnode;
1823 		VREF(rootvnode);
1824 	}
1825 	if (!curthread->td_proc->p_fd->fd_rdir) {
1826 		curthread->td_proc->p_fd->fd_rdir = rootvnode;
1827 		VREF(rootvnode);
1828 	}
1829 	if (!curthread->td_proc->p_fd->fd_jdir) {
1830 		curthread->td_proc->p_fd->fd_jdir = rootvnode;
1831 		VREF(rootvnode);
1832 	}
1833 
1834  again:
1835 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, be_lun->dev_path, curthread);
1836 	error = vn_open(&nd, &flags, 0, NULL);
1837 	if (error) {
1838 		/*
1839 		 * This is the only reasonable guess we can make as far as
1840 		 * path if the user doesn't give us a fully qualified path.
1841 		 * If they want to specify a file, they need to specify the
1842 		 * full path.
1843 		 */
1844 		if (be_lun->dev_path[0] != '/') {
1845 			char *dev_path = "/dev/";
1846 			char *dev_name;
1847 
1848 			/* Try adding device path at beginning of name */
1849 			dev_name = malloc(strlen(be_lun->dev_path)
1850 					+ strlen(dev_path) + 1,
1851 					  M_CTLBLK, M_WAITOK);
1852 			if (dev_name) {
1853 				sprintf(dev_name, "%s%s", dev_path,
1854 					be_lun->dev_path);
1855 				free(be_lun->dev_path, M_CTLBLK);
1856 				be_lun->dev_path = dev_name;
1857 				goto again;
1858 			}
1859 		}
1860 		snprintf(req->error_str, sizeof(req->error_str),
1861 			 "%s: error opening %s", __func__, be_lun->dev_path);
1862 		return (error);
1863 	}
1864 
1865 	NDFREE(&nd, NDF_ONLY_PNBUF);
1866 
1867 	be_lun->vn = nd.ni_vp;
1868 
1869 	/* We only support disks and files. */
1870 	if (vn_isdisk(be_lun->vn, &error)) {
1871 		error = ctl_be_block_open_dev(be_lun, req);
1872 	} else if (be_lun->vn->v_type == VREG) {
1873 		error = ctl_be_block_open_file(be_lun, req);
1874 	} else {
1875 		error = EINVAL;
1876 		snprintf(req->error_str, sizeof(req->error_str),
1877 			 "%s is not a disk or plain file", be_lun->dev_path);
1878 	}
1879 	VOP_UNLOCK(be_lun->vn, 0);
1880 
1881 	if (error != 0) {
1882 		ctl_be_block_close(be_lun);
1883 		return (error);
1884 	}
1885 
1886 	be_lun->blocksize_shift = fls(be_lun->blocksize) - 1;
1887 	be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
1888 
1889 	return (0);
1890 }
1891 
1892 static int
1893 ctl_be_block_create(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
1894 {
1895 	struct ctl_be_block_lun *be_lun;
1896 	struct ctl_lun_create_params *params;
1897 	char num_thread_str[16];
1898 	char tmpstr[32];
1899 	char *value;
1900 	int retval, num_threads, unmap;
1901 	int tmp_num_threads;
1902 
1903 	params = &req->reqdata.create;
1904 	retval = 0;
1905 
1906 	num_threads = cbb_num_threads;
1907 
1908 	be_lun = malloc(sizeof(*be_lun), M_CTLBLK, M_ZERO | M_WAITOK);
1909 
1910 	be_lun->softc = softc;
1911 	STAILQ_INIT(&be_lun->input_queue);
1912 	STAILQ_INIT(&be_lun->config_write_queue);
1913 	STAILQ_INIT(&be_lun->datamove_queue);
1914 	sprintf(be_lun->lunname, "cblk%d", softc->num_luns);
1915 	mtx_init(&be_lun->io_lock, "cblk io lock", NULL, MTX_DEF);
1916 	mtx_init(&be_lun->queue_lock, "cblk queue lock", NULL, MTX_DEF);
1917 	ctl_init_opts(&be_lun->ctl_be_lun.options,
1918 	    req->num_be_args, req->kern_be_args);
1919 
1920 	be_lun->lun_zone = uma_zcreate(be_lun->lunname, CTLBLK_MAX_SEG,
1921 	    NULL, NULL, NULL, NULL, /*align*/ 0, /*flags*/0);
1922 
1923 	if (be_lun->lun_zone == NULL) {
1924 		snprintf(req->error_str, sizeof(req->error_str),
1925 			 "%s: error allocating UMA zone", __func__);
1926 		goto bailout_error;
1927 	}
1928 
1929 	if (params->flags & CTL_LUN_FLAG_DEV_TYPE)
1930 		be_lun->ctl_be_lun.lun_type = params->device_type;
1931 	else
1932 		be_lun->ctl_be_lun.lun_type = T_DIRECT;
1933 
1934 	if (be_lun->ctl_be_lun.lun_type == T_DIRECT) {
1935 		value = ctl_get_opt(&be_lun->ctl_be_lun.options, "file");
1936 		if (value == NULL) {
1937 			snprintf(req->error_str, sizeof(req->error_str),
1938 				 "%s: no file argument specified", __func__);
1939 			goto bailout_error;
1940 		}
1941 		be_lun->dev_path = strdup(value, M_CTLBLK);
1942 
1943 		retval = ctl_be_block_open(softc, be_lun, req);
1944 		if (retval != 0) {
1945 			retval = 0;
1946 			goto bailout_error;
1947 		}
1948 
1949 		/*
1950 		 * Tell the user the size of the file/device.
1951 		 */
1952 		params->lun_size_bytes = be_lun->size_bytes;
1953 
1954 		/*
1955 		 * The maximum LBA is the size - 1.
1956 		 */
1957 		be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
1958 	} else {
1959 		/*
1960 		 * For processor devices, we don't have any size.
1961 		 */
1962 		be_lun->blocksize = 0;
1963 		be_lun->pblockexp = 0;
1964 		be_lun->pblockoff = 0;
1965 		be_lun->size_blocks = 0;
1966 		be_lun->size_bytes = 0;
1967 		be_lun->ctl_be_lun.maxlba = 0;
1968 		params->lun_size_bytes = 0;
1969 
1970 		/*
1971 		 * Default to just 1 thread for processor devices.
1972 		 */
1973 		num_threads = 1;
1974 	}
1975 
1976 	/*
1977 	 * XXX This searching loop might be refactored to be combined with
1978 	 * the loop above,
1979 	 */
1980 	value = ctl_get_opt(&be_lun->ctl_be_lun.options, "num_threads");
1981 	if (value != NULL) {
1982 		tmp_num_threads = strtol(value, NULL, 0);
1983 
1984 		/*
1985 		 * We don't let the user specify less than one
1986 		 * thread, but hope he's clueful enough not to
1987 		 * specify 1000 threads.
1988 		 */
1989 		if (tmp_num_threads < 1) {
1990 			snprintf(req->error_str, sizeof(req->error_str),
1991 				 "%s: invalid number of threads %s",
1992 			         __func__, num_thread_str);
1993 			goto bailout_error;
1994 		}
1995 		num_threads = tmp_num_threads;
1996 	}
1997 	unmap = 0;
1998 	value = ctl_get_opt(&be_lun->ctl_be_lun.options, "unmap");
1999 	if (value != NULL && strcmp(value, "on") == 0)
2000 		unmap = 1;
2001 
2002 	be_lun->flags = CTL_BE_BLOCK_LUN_UNCONFIGURED;
2003 	be_lun->ctl_be_lun.flags = CTL_LUN_FLAG_PRIMARY;
2004 	if (unmap)
2005 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_UNMAP;
2006 	be_lun->ctl_be_lun.be_lun = be_lun;
2007 	be_lun->ctl_be_lun.blocksize = be_lun->blocksize;
2008 	be_lun->ctl_be_lun.pblockexp = be_lun->pblockexp;
2009 	be_lun->ctl_be_lun.pblockoff = be_lun->pblockoff;
2010 	/* Tell the user the blocksize we ended up using */
2011 	params->blocksize_bytes = be_lun->blocksize;
2012 	if (params->flags & CTL_LUN_FLAG_ID_REQ) {
2013 		be_lun->ctl_be_lun.req_lun_id = params->req_lun_id;
2014 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_ID_REQ;
2015 	} else
2016 		be_lun->ctl_be_lun.req_lun_id = 0;
2017 
2018 	be_lun->ctl_be_lun.lun_shutdown = ctl_be_block_lun_shutdown;
2019 	be_lun->ctl_be_lun.lun_config_status =
2020 		ctl_be_block_lun_config_status;
2021 	be_lun->ctl_be_lun.be = &ctl_be_block_driver;
2022 
2023 	if ((params->flags & CTL_LUN_FLAG_SERIAL_NUM) == 0) {
2024 		snprintf(tmpstr, sizeof(tmpstr), "MYSERIAL%4d",
2025 			 softc->num_luns);
2026 		strncpy((char *)be_lun->ctl_be_lun.serial_num, tmpstr,
2027 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
2028 			sizeof(tmpstr)));
2029 
2030 		/* Tell the user what we used for a serial number */
2031 		strncpy((char *)params->serial_num, tmpstr,
2032 			ctl_min(sizeof(params->serial_num), sizeof(tmpstr)));
2033 	} else {
2034 		strncpy((char *)be_lun->ctl_be_lun.serial_num,
2035 			params->serial_num,
2036 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
2037 			sizeof(params->serial_num)));
2038 	}
2039 	if ((params->flags & CTL_LUN_FLAG_DEVID) == 0) {
2040 		snprintf(tmpstr, sizeof(tmpstr), "MYDEVID%4d", softc->num_luns);
2041 		strncpy((char *)be_lun->ctl_be_lun.device_id, tmpstr,
2042 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
2043 			sizeof(tmpstr)));
2044 
2045 		/* Tell the user what we used for a device ID */
2046 		strncpy((char *)params->device_id, tmpstr,
2047 			ctl_min(sizeof(params->device_id), sizeof(tmpstr)));
2048 	} else {
2049 		strncpy((char *)be_lun->ctl_be_lun.device_id,
2050 			params->device_id,
2051 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
2052 				sizeof(params->device_id)));
2053 	}
2054 
2055 	TASK_INIT(&be_lun->io_task, /*priority*/0, ctl_be_block_worker, be_lun);
2056 
2057 	be_lun->io_taskqueue = taskqueue_create(be_lun->lunname, M_WAITOK,
2058 	    taskqueue_thread_enqueue, /*context*/&be_lun->io_taskqueue);
2059 
2060 	if (be_lun->io_taskqueue == NULL) {
2061 		snprintf(req->error_str, sizeof(req->error_str),
2062 			 "%s: Unable to create taskqueue", __func__);
2063 		goto bailout_error;
2064 	}
2065 
2066 	/*
2067 	 * Note that we start the same number of threads by default for
2068 	 * both the file case and the block device case.  For the file
2069 	 * case, we need multiple threads to allow concurrency, because the
2070 	 * vnode interface is designed to be a blocking interface.  For the
2071 	 * block device case, ZFS zvols at least will block the caller's
2072 	 * context in many instances, and so we need multiple threads to
2073 	 * overcome that problem.  Other block devices don't need as many
2074 	 * threads, but they shouldn't cause too many problems.
2075 	 *
2076 	 * If the user wants to just have a single thread for a block
2077 	 * device, he can specify that when the LUN is created, or change
2078 	 * the tunable/sysctl to alter the default number of threads.
2079 	 */
2080 	retval = taskqueue_start_threads(&be_lun->io_taskqueue,
2081 					 /*num threads*/num_threads,
2082 					 /*priority*/PWAIT,
2083 					 /*thread name*/
2084 					 "%s taskq", be_lun->lunname);
2085 
2086 	if (retval != 0)
2087 		goto bailout_error;
2088 
2089 	be_lun->num_threads = num_threads;
2090 
2091 	mtx_lock(&softc->lock);
2092 	softc->num_luns++;
2093 	STAILQ_INSERT_TAIL(&softc->lun_list, be_lun, links);
2094 
2095 	mtx_unlock(&softc->lock);
2096 
2097 	retval = ctl_add_lun(&be_lun->ctl_be_lun);
2098 	if (retval != 0) {
2099 		mtx_lock(&softc->lock);
2100 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2101 			      links);
2102 		softc->num_luns--;
2103 		mtx_unlock(&softc->lock);
2104 		snprintf(req->error_str, sizeof(req->error_str),
2105 			 "%s: ctl_add_lun() returned error %d, see dmesg for "
2106 			"details", __func__, retval);
2107 		retval = 0;
2108 		goto bailout_error;
2109 	}
2110 
2111 	mtx_lock(&softc->lock);
2112 
2113 	/*
2114 	 * Tell the config_status routine that we're waiting so it won't
2115 	 * clean up the LUN in the event of an error.
2116 	 */
2117 	be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
2118 
2119 	while (be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) {
2120 		retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
2121 		if (retval == EINTR)
2122 			break;
2123 	}
2124 	be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
2125 
2126 	if (be_lun->flags & CTL_BE_BLOCK_LUN_CONFIG_ERR) {
2127 		snprintf(req->error_str, sizeof(req->error_str),
2128 			 "%s: LUN configuration error, see dmesg for details",
2129 			 __func__);
2130 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2131 			      links);
2132 		softc->num_luns--;
2133 		mtx_unlock(&softc->lock);
2134 		goto bailout_error;
2135 	} else {
2136 		params->req_lun_id = be_lun->ctl_be_lun.lun_id;
2137 	}
2138 
2139 	mtx_unlock(&softc->lock);
2140 
2141 	be_lun->disk_stats = devstat_new_entry("cbb", params->req_lun_id,
2142 					       be_lun->blocksize,
2143 					       DEVSTAT_ALL_SUPPORTED,
2144 					       be_lun->ctl_be_lun.lun_type
2145 					       | DEVSTAT_TYPE_IF_OTHER,
2146 					       DEVSTAT_PRIORITY_OTHER);
2147 
2148 
2149 	req->status = CTL_LUN_OK;
2150 
2151 	return (retval);
2152 
2153 bailout_error:
2154 	req->status = CTL_LUN_ERROR;
2155 
2156 	if (be_lun->io_taskqueue != NULL)
2157 		taskqueue_free(be_lun->io_taskqueue);
2158 	ctl_be_block_close(be_lun);
2159 	if (be_lun->dev_path != NULL)
2160 		free(be_lun->dev_path, M_CTLBLK);
2161 	if (be_lun->lun_zone != NULL)
2162 		uma_zdestroy(be_lun->lun_zone);
2163 	ctl_free_opts(&be_lun->ctl_be_lun.options);
2164 	mtx_destroy(&be_lun->queue_lock);
2165 	mtx_destroy(&be_lun->io_lock);
2166 	free(be_lun, M_CTLBLK);
2167 
2168 	return (retval);
2169 }
2170 
2171 static int
2172 ctl_be_block_rm(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2173 {
2174 	struct ctl_lun_rm_params *params;
2175 	struct ctl_be_block_lun *be_lun;
2176 	int retval;
2177 
2178 	params = &req->reqdata.rm;
2179 
2180 	mtx_lock(&softc->lock);
2181 
2182 	be_lun = NULL;
2183 
2184 	STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2185 		if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2186 			break;
2187 	}
2188 	mtx_unlock(&softc->lock);
2189 
2190 	if (be_lun == NULL) {
2191 		snprintf(req->error_str, sizeof(req->error_str),
2192 			 "%s: LUN %u is not managed by the block backend",
2193 			 __func__, params->lun_id);
2194 		goto bailout_error;
2195 	}
2196 
2197 	retval = ctl_disable_lun(&be_lun->ctl_be_lun);
2198 
2199 	if (retval != 0) {
2200 		snprintf(req->error_str, sizeof(req->error_str),
2201 			 "%s: error %d returned from ctl_disable_lun() for "
2202 			 "LUN %d", __func__, retval, params->lun_id);
2203 		goto bailout_error;
2204 
2205 	}
2206 
2207 	retval = ctl_invalidate_lun(&be_lun->ctl_be_lun);
2208 	if (retval != 0) {
2209 		snprintf(req->error_str, sizeof(req->error_str),
2210 			 "%s: error %d returned from ctl_invalidate_lun() for "
2211 			 "LUN %d", __func__, retval, params->lun_id);
2212 		goto bailout_error;
2213 	}
2214 
2215 	mtx_lock(&softc->lock);
2216 
2217 	be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
2218 
2219 	while ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
2220                 retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
2221                 if (retval == EINTR)
2222                         break;
2223         }
2224 
2225 	be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
2226 
2227 	if ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
2228 		snprintf(req->error_str, sizeof(req->error_str),
2229 			 "%s: interrupted waiting for LUN to be freed",
2230 			 __func__);
2231 		mtx_unlock(&softc->lock);
2232 		goto bailout_error;
2233 	}
2234 
2235 	STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun, links);
2236 
2237 	softc->num_luns--;
2238 	mtx_unlock(&softc->lock);
2239 
2240 	taskqueue_drain(be_lun->io_taskqueue, &be_lun->io_task);
2241 
2242 	taskqueue_free(be_lun->io_taskqueue);
2243 
2244 	ctl_be_block_close(be_lun);
2245 
2246 	if (be_lun->disk_stats != NULL)
2247 		devstat_remove_entry(be_lun->disk_stats);
2248 
2249 	uma_zdestroy(be_lun->lun_zone);
2250 
2251 	ctl_free_opts(&be_lun->ctl_be_lun.options);
2252 	free(be_lun->dev_path, M_CTLBLK);
2253 	mtx_destroy(&be_lun->queue_lock);
2254 	mtx_destroy(&be_lun->io_lock);
2255 	free(be_lun, M_CTLBLK);
2256 
2257 	req->status = CTL_LUN_OK;
2258 
2259 	return (0);
2260 
2261 bailout_error:
2262 
2263 	req->status = CTL_LUN_ERROR;
2264 
2265 	return (0);
2266 }
2267 
2268 static int
2269 ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
2270 			 struct ctl_lun_req *req)
2271 {
2272 	struct vattr vattr;
2273 	int error;
2274 	struct ctl_lun_modify_params *params;
2275 
2276 	params = &req->reqdata.modify;
2277 
2278 	if (params->lun_size_bytes != 0) {
2279 		be_lun->size_bytes = params->lun_size_bytes;
2280 	} else  {
2281 		error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
2282 		if (error != 0) {
2283 			snprintf(req->error_str, sizeof(req->error_str),
2284 				 "error calling VOP_GETATTR() for file %s",
2285 				 be_lun->dev_path);
2286 			return (error);
2287 		}
2288 
2289 		be_lun->size_bytes = vattr.va_size;
2290 	}
2291 
2292 	return (0);
2293 }
2294 
2295 static int
2296 ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
2297 			struct ctl_lun_req *req)
2298 {
2299 	struct cdev *dev;
2300 	struct cdevsw *devsw;
2301 	int error;
2302 	struct ctl_lun_modify_params *params;
2303 	uint64_t size_bytes;
2304 
2305 	params = &req->reqdata.modify;
2306 
2307 	dev = be_lun->vn->v_rdev;
2308 	devsw = dev->si_devsw;
2309 	if (!devsw->d_ioctl) {
2310 		snprintf(req->error_str, sizeof(req->error_str),
2311 			 "%s: no d_ioctl for device %s!", __func__,
2312 			 be_lun->dev_path);
2313 		return (ENODEV);
2314 	}
2315 
2316 	error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
2317 			       (caddr_t)&size_bytes, FREAD,
2318 			       curthread);
2319 	if (error) {
2320 		snprintf(req->error_str, sizeof(req->error_str),
2321 			 "%s: error %d returned for DIOCGMEDIASIZE ioctl "
2322 			 "on %s!", __func__, error, be_lun->dev_path);
2323 		return (error);
2324 	}
2325 
2326 	if (params->lun_size_bytes != 0) {
2327 		if (params->lun_size_bytes > size_bytes) {
2328 			snprintf(req->error_str, sizeof(req->error_str),
2329 				 "%s: requested LUN size %ju > backing device "
2330 				 "size %ju", __func__,
2331 				 (uintmax_t)params->lun_size_bytes,
2332 				 (uintmax_t)size_bytes);
2333 			return (EINVAL);
2334 		}
2335 
2336 		be_lun->size_bytes = params->lun_size_bytes;
2337 	} else {
2338 		be_lun->size_bytes = size_bytes;
2339 	}
2340 
2341 	return (0);
2342 }
2343 
2344 static int
2345 ctl_be_block_modify(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2346 {
2347 	struct ctl_lun_modify_params *params;
2348 	struct ctl_be_block_lun *be_lun;
2349 	int error;
2350 
2351 	params = &req->reqdata.modify;
2352 
2353 	mtx_lock(&softc->lock);
2354 
2355 	be_lun = NULL;
2356 
2357 	STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2358 		if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2359 			break;
2360 	}
2361 	mtx_unlock(&softc->lock);
2362 
2363 	if (be_lun == NULL) {
2364 		snprintf(req->error_str, sizeof(req->error_str),
2365 			 "%s: LUN %u is not managed by the block backend",
2366 			 __func__, params->lun_id);
2367 		goto bailout_error;
2368 	}
2369 
2370 	if (params->lun_size_bytes != 0) {
2371 		if (params->lun_size_bytes < be_lun->blocksize) {
2372 			snprintf(req->error_str, sizeof(req->error_str),
2373 				"%s: LUN size %ju < blocksize %u", __func__,
2374 				params->lun_size_bytes, be_lun->blocksize);
2375 			goto bailout_error;
2376 		}
2377 	}
2378 
2379 	vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
2380 
2381 	if (be_lun->vn->v_type == VREG)
2382 		error = ctl_be_block_modify_file(be_lun, req);
2383 	else
2384 		error = ctl_be_block_modify_dev(be_lun, req);
2385 
2386 	VOP_UNLOCK(be_lun->vn, 0);
2387 
2388 	if (error != 0)
2389 		goto bailout_error;
2390 
2391 	be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
2392 
2393 	/*
2394 	 * The maximum LBA is the size - 1.
2395 	 *
2396 	 * XXX: Note that this field is being updated without locking,
2397 	 * 	which might cause problems on 32-bit architectures.
2398 	 */
2399 	be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
2400 	ctl_lun_capacity_changed(&be_lun->ctl_be_lun);
2401 
2402 	/* Tell the user the exact size we ended up using */
2403 	params->lun_size_bytes = be_lun->size_bytes;
2404 
2405 	req->status = CTL_LUN_OK;
2406 
2407 	return (0);
2408 
2409 bailout_error:
2410 	req->status = CTL_LUN_ERROR;
2411 
2412 	return (0);
2413 }
2414 
2415 static void
2416 ctl_be_block_lun_shutdown(void *be_lun)
2417 {
2418 	struct ctl_be_block_lun *lun;
2419 	struct ctl_be_block_softc *softc;
2420 
2421 	lun = (struct ctl_be_block_lun *)be_lun;
2422 
2423 	softc = lun->softc;
2424 
2425 	mtx_lock(&softc->lock);
2426 	lun->flags |= CTL_BE_BLOCK_LUN_UNCONFIGURED;
2427 	if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2428 		wakeup(lun);
2429 	mtx_unlock(&softc->lock);
2430 
2431 }
2432 
2433 static void
2434 ctl_be_block_lun_config_status(void *be_lun, ctl_lun_config_status status)
2435 {
2436 	struct ctl_be_block_lun *lun;
2437 	struct ctl_be_block_softc *softc;
2438 
2439 	lun = (struct ctl_be_block_lun *)be_lun;
2440 	softc = lun->softc;
2441 
2442 	if (status == CTL_LUN_CONFIG_OK) {
2443 		mtx_lock(&softc->lock);
2444 		lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2445 		if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2446 			wakeup(lun);
2447 		mtx_unlock(&softc->lock);
2448 
2449 		/*
2450 		 * We successfully added the LUN, attempt to enable it.
2451 		 */
2452 		if (ctl_enable_lun(&lun->ctl_be_lun) != 0) {
2453 			printf("%s: ctl_enable_lun() failed!\n", __func__);
2454 			if (ctl_invalidate_lun(&lun->ctl_be_lun) != 0) {
2455 				printf("%s: ctl_invalidate_lun() failed!\n",
2456 				       __func__);
2457 			}
2458 		}
2459 
2460 		return;
2461 	}
2462 
2463 
2464 	mtx_lock(&softc->lock);
2465 	lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2466 	lun->flags |= CTL_BE_BLOCK_LUN_CONFIG_ERR;
2467 	wakeup(lun);
2468 	mtx_unlock(&softc->lock);
2469 }
2470 
2471 
2472 static int
2473 ctl_be_block_config_write(union ctl_io *io)
2474 {
2475 	struct ctl_be_block_lun *be_lun;
2476 	struct ctl_be_lun *ctl_be_lun;
2477 	int retval;
2478 
2479 	retval = 0;
2480 
2481 	DPRINTF("entered\n");
2482 
2483 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
2484 		CTL_PRIV_BACKEND_LUN].ptr;
2485 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
2486 
2487 	switch (io->scsiio.cdb[0]) {
2488 	case SYNCHRONIZE_CACHE:
2489 	case SYNCHRONIZE_CACHE_16:
2490 	case WRITE_SAME_10:
2491 	case WRITE_SAME_16:
2492 	case UNMAP:
2493 		/*
2494 		 * The upper level CTL code will filter out any CDBs with
2495 		 * the immediate bit set and return the proper error.
2496 		 *
2497 		 * We don't really need to worry about what LBA range the
2498 		 * user asked to be synced out.  When they issue a sync
2499 		 * cache command, we'll sync out the whole thing.
2500 		 */
2501 		mtx_lock(&be_lun->queue_lock);
2502 		STAILQ_INSERT_TAIL(&be_lun->config_write_queue, &io->io_hdr,
2503 				   links);
2504 		mtx_unlock(&be_lun->queue_lock);
2505 		taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
2506 		break;
2507 	case START_STOP_UNIT: {
2508 		struct scsi_start_stop_unit *cdb;
2509 
2510 		cdb = (struct scsi_start_stop_unit *)io->scsiio.cdb;
2511 
2512 		if (cdb->how & SSS_START)
2513 			retval = ctl_start_lun(ctl_be_lun);
2514 		else {
2515 			retval = ctl_stop_lun(ctl_be_lun);
2516 			/*
2517 			 * XXX KDM Copan-specific offline behavior.
2518 			 * Figure out a reasonable way to port this?
2519 			 */
2520 #ifdef NEEDTOPORT
2521 			if ((retval == 0)
2522 			 && (cdb->byte2 & SSS_ONOFFLINE))
2523 				retval = ctl_lun_offline(ctl_be_lun);
2524 #endif
2525 		}
2526 
2527 		/*
2528 		 * In general, the above routines should not fail.  They
2529 		 * just set state for the LUN.  So we've got something
2530 		 * pretty wrong here if we can't start or stop the LUN.
2531 		 */
2532 		if (retval != 0) {
2533 			ctl_set_internal_failure(&io->scsiio,
2534 						 /*sks_valid*/ 1,
2535 						 /*retry_count*/ 0xf051);
2536 			retval = CTL_RETVAL_COMPLETE;
2537 		} else {
2538 			ctl_set_success(&io->scsiio);
2539 		}
2540 		ctl_config_write_done(io);
2541 		break;
2542 	}
2543 	default:
2544 		ctl_set_invalid_opcode(&io->scsiio);
2545 		ctl_config_write_done(io);
2546 		retval = CTL_RETVAL_COMPLETE;
2547 		break;
2548 	}
2549 
2550 	return (retval);
2551 
2552 }
2553 
2554 static int
2555 ctl_be_block_config_read(union ctl_io *io)
2556 {
2557 	return (0);
2558 }
2559 
2560 static int
2561 ctl_be_block_lun_info(void *be_lun, struct sbuf *sb)
2562 {
2563 	struct ctl_be_block_lun *lun;
2564 	int retval;
2565 
2566 	lun = (struct ctl_be_block_lun *)be_lun;
2567 	retval = 0;
2568 
2569 	retval = sbuf_printf(sb, "\t<num_threads>");
2570 
2571 	if (retval != 0)
2572 		goto bailout;
2573 
2574 	retval = sbuf_printf(sb, "%d", lun->num_threads);
2575 
2576 	if (retval != 0)
2577 		goto bailout;
2578 
2579 	retval = sbuf_printf(sb, "</num_threads>\n");
2580 
2581 bailout:
2582 
2583 	return (retval);
2584 }
2585 
2586 int
2587 ctl_be_block_init(void)
2588 {
2589 	struct ctl_be_block_softc *softc;
2590 	int retval;
2591 
2592 	softc = &backend_block_softc;
2593 	retval = 0;
2594 
2595 	mtx_init(&softc->lock, "ctlblock", NULL, MTX_DEF);
2596 	beio_zone = uma_zcreate("beio", sizeof(struct ctl_be_block_io),
2597 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
2598 	STAILQ_INIT(&softc->disk_list);
2599 	STAILQ_INIT(&softc->lun_list);
2600 
2601 	return (retval);
2602 }
2603