xref: /freebsd/sys/cam/ctl/ctl_backend_block.c (revision 5bd73b51076b5cb5a2c9810f76c1d7ed20c4460e)
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_lun_create_params params;
155 	struct ctl_block_disk *disk;
156 	char lunname[32];
157 	char *dev_path;
158 	ctl_be_block_type dev_type;
159 	struct vnode *vn;
160 	union ctl_be_block_bedata backend;
161 	cbb_dispatch_t dispatch;
162 	cbb_dispatch_t lun_flush;
163 	cbb_dispatch_t unmap;
164 	uma_zone_t lun_zone;
165 	uint64_t size_blocks;
166 	uint64_t size_bytes;
167 	uint32_t blocksize;
168 	int blocksize_shift;
169 	uint16_t pblockexp;
170 	uint16_t pblockoff;
171 	struct ctl_be_block_softc *softc;
172 	struct devstat *disk_stats;
173 	ctl_be_block_lun_flags flags;
174 	STAILQ_ENTRY(ctl_be_block_lun) links;
175 	struct ctl_be_lun ctl_be_lun;
176 	struct taskqueue *io_taskqueue;
177 	struct task io_task;
178 	int num_threads;
179 	STAILQ_HEAD(, ctl_io_hdr) input_queue;
180 	STAILQ_HEAD(, ctl_io_hdr) config_write_queue;
181 	STAILQ_HEAD(, ctl_io_hdr) datamove_queue;
182 	struct mtx_padalign io_lock;
183 	struct mtx_padalign queue_lock;
184 };
185 
186 /*
187  * Overall softc structure for the block backend module.
188  */
189 struct ctl_be_block_softc {
190 	struct mtx			 lock;
191 	int				 num_disks;
192 	STAILQ_HEAD(, ctl_block_disk)	 disk_list;
193 	int				 num_luns;
194 	STAILQ_HEAD(, ctl_be_block_lun)	 lun_list;
195 };
196 
197 static struct ctl_be_block_softc backend_block_softc;
198 
199 /*
200  * Per-I/O information.
201  */
202 struct ctl_be_block_io {
203 	union ctl_io			*io;
204 	struct ctl_sg_entry		sg_segs[CTLBLK_MAX_SEGS];
205 	struct iovec			xiovecs[CTLBLK_MAX_SEGS];
206 	int				bio_cmd;
207 	int				num_segs;
208 	int				num_bios_sent;
209 	int				num_bios_done;
210 	int				send_complete;
211 	int				num_errors;
212 	struct bintime			ds_t0;
213 	devstat_tag_type		ds_tag_type;
214 	devstat_trans_flags		ds_trans_type;
215 	uint64_t			io_len;
216 	uint64_t			io_offset;
217 	struct ctl_be_block_softc	*softc;
218 	struct ctl_be_block_lun		*lun;
219 	void (*beio_cont)(struct ctl_be_block_io *beio); /* to continue processing */
220 };
221 
222 static int cbb_num_threads = 14;
223 SYSCTL_NODE(_kern_cam_ctl, OID_AUTO, block, CTLFLAG_RD, 0,
224 	    "CAM Target Layer Block Backend");
225 SYSCTL_INT(_kern_cam_ctl_block, OID_AUTO, num_threads, CTLFLAG_RWTUN,
226            &cbb_num_threads, 0, "Number of threads per backing file");
227 
228 static struct ctl_be_block_io *ctl_alloc_beio(struct ctl_be_block_softc *softc);
229 static void ctl_free_beio(struct ctl_be_block_io *beio);
230 static void ctl_complete_beio(struct ctl_be_block_io *beio);
231 static int ctl_be_block_move_done(union ctl_io *io);
232 static void ctl_be_block_biodone(struct bio *bio);
233 static void ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
234 				    struct ctl_be_block_io *beio);
235 static void ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
236 				       struct ctl_be_block_io *beio);
237 static void ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
238 				   struct ctl_be_block_io *beio);
239 static void ctl_be_block_unmap_dev(struct ctl_be_block_lun *be_lun,
240 				   struct ctl_be_block_io *beio);
241 static void ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
242 				      struct ctl_be_block_io *beio);
243 static void ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
244 				    union ctl_io *io);
245 static void ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
246 				  union ctl_io *io);
247 static void ctl_be_block_worker(void *context, int pending);
248 static int ctl_be_block_submit(union ctl_io *io);
249 static int ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
250 				   int flag, struct thread *td);
251 static int ctl_be_block_open_file(struct ctl_be_block_lun *be_lun,
252 				  struct ctl_lun_req *req);
253 static int ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun,
254 				 struct ctl_lun_req *req);
255 static int ctl_be_block_close(struct ctl_be_block_lun *be_lun);
256 static int ctl_be_block_open(struct ctl_be_block_softc *softc,
257 			     struct ctl_be_block_lun *be_lun,
258 			     struct ctl_lun_req *req);
259 static int ctl_be_block_create(struct ctl_be_block_softc *softc,
260 			       struct ctl_lun_req *req);
261 static int ctl_be_block_rm(struct ctl_be_block_softc *softc,
262 			   struct ctl_lun_req *req);
263 static int ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
264 				  struct ctl_lun_req *req);
265 static int ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
266 				 struct ctl_lun_req *req);
267 static int ctl_be_block_modify(struct ctl_be_block_softc *softc,
268 			   struct ctl_lun_req *req);
269 static void ctl_be_block_lun_shutdown(void *be_lun);
270 static void ctl_be_block_lun_config_status(void *be_lun,
271 					   ctl_lun_config_status status);
272 static int ctl_be_block_config_write(union ctl_io *io);
273 static int ctl_be_block_config_read(union ctl_io *io);
274 static int ctl_be_block_lun_info(void *be_lun, struct sbuf *sb);
275 int ctl_be_block_init(void);
276 
277 static struct ctl_backend_driver ctl_be_block_driver =
278 {
279 	.name = "block",
280 	.flags = CTL_BE_FLAG_HAS_CONFIG,
281 	.init = ctl_be_block_init,
282 	.data_submit = ctl_be_block_submit,
283 	.data_move_done = ctl_be_block_move_done,
284 	.config_read = ctl_be_block_config_read,
285 	.config_write = ctl_be_block_config_write,
286 	.ioctl = ctl_be_block_ioctl,
287 	.lun_info = ctl_be_block_lun_info
288 };
289 
290 MALLOC_DEFINE(M_CTLBLK, "ctlblk", "Memory used for CTL block backend");
291 CTL_BACKEND_DECLARE(cbb, ctl_be_block_driver);
292 
293 static uma_zone_t beio_zone;
294 
295 static struct ctl_be_block_io *
296 ctl_alloc_beio(struct ctl_be_block_softc *softc)
297 {
298 	struct ctl_be_block_io *beio;
299 
300 	beio = uma_zalloc(beio_zone, M_WAITOK | M_ZERO);
301 	beio->softc = softc;
302 	return (beio);
303 }
304 
305 static void
306 ctl_free_beio(struct ctl_be_block_io *beio)
307 {
308 	int duplicate_free;
309 	int i;
310 
311 	duplicate_free = 0;
312 
313 	for (i = 0; i < beio->num_segs; i++) {
314 		if (beio->sg_segs[i].addr == NULL)
315 			duplicate_free++;
316 
317 		uma_zfree(beio->lun->lun_zone, beio->sg_segs[i].addr);
318 		beio->sg_segs[i].addr = NULL;
319 
320 		/* For compare we had two equal S/G lists. */
321 		if (ARGS(beio->io)->flags & CTL_LLF_COMPARE) {
322 			uma_zfree(beio->lun->lun_zone,
323 			    beio->sg_segs[i + CTLBLK_HALF_SEGS].addr);
324 			beio->sg_segs[i + CTLBLK_HALF_SEGS].addr = NULL;
325 		}
326 	}
327 
328 	if (duplicate_free > 0) {
329 		printf("%s: %d duplicate frees out of %d segments\n", __func__,
330 		       duplicate_free, beio->num_segs);
331 	}
332 
333 	uma_zfree(beio_zone, beio);
334 }
335 
336 static void
337 ctl_complete_beio(struct ctl_be_block_io *beio)
338 {
339 	union ctl_io *io = beio->io;
340 
341 	if (beio->beio_cont != NULL) {
342 		beio->beio_cont(beio);
343 	} else {
344 		ctl_free_beio(beio);
345 		ctl_data_submit_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 	struct ctl_lba_len_flags *lbalen;
355 #ifdef CTL_TIME_IO
356 	struct bintime cur_bt;
357 #endif
358 	int i;
359 
360 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
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 	io->scsiio.kern_rel_offset += io->scsiio.kern_data_len;
372 
373 	/*
374 	 * We set status at this point for read commands, and write
375 	 * commands with errors.
376 	 */
377 	if ((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 		lbalen = ARGS(beio->io);
381 		if (lbalen->flags & CTL_LLF_READ) {
382 			ctl_set_success(&io->scsiio);
383 		} else if (lbalen->flags & CTL_LLF_COMPARE) {
384 			/* We have two data blocks ready for comparison. */
385 			for (i = 0; i < beio->num_segs; i++) {
386 				if (memcmp(beio->sg_segs[i].addr,
387 				    beio->sg_segs[i + CTLBLK_HALF_SEGS].addr,
388 				    beio->sg_segs[i].len) != 0)
389 					break;
390 			}
391 			if (i < beio->num_segs)
392 				ctl_set_sense(&io->scsiio,
393 				    /*current_error*/ 1,
394 				    /*sense_key*/ SSD_KEY_MISCOMPARE,
395 				    /*asc*/ 0x1D,
396 				    /*ascq*/ 0x00,
397 				    SSD_ELEM_NONE);
398 			else
399 				ctl_set_success(&io->scsiio);
400 		}
401 	}
402 	else if ((io->io_hdr.port_status != 0)
403 	      && ((io->io_hdr.flags & CTL_FLAG_ABORT) == 0)
404 	      && ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_STATUS_NONE)) {
405 		/*
406 		 * For hardware error sense keys, the sense key
407 		 * specific value is defined to be a retry count,
408 		 * but we use it to pass back an internal FETD
409 		 * error code.  XXX KDM  Hopefully the FETD is only
410 		 * using 16 bits for an error code, since that's
411 		 * all the space we have in the sks field.
412 		 */
413 		ctl_set_internal_failure(&io->scsiio,
414 					 /*sks_valid*/ 1,
415 					 /*retry_count*/
416 					 io->io_hdr.port_status);
417 	}
418 
419 	/*
420 	 * If this is a read, or a write with errors, it is done.
421 	 */
422 	if ((beio->bio_cmd == BIO_READ)
423 	 || ((io->io_hdr.flags & CTL_FLAG_ABORT) != 0)
424 	 || ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE)) {
425 		ctl_complete_beio(beio);
426 		return (0);
427 	}
428 
429 	/*
430 	 * At this point, we have a write and the DMA completed
431 	 * successfully.  We now have to queue it to the task queue to
432 	 * execute the backend I/O.  That is because we do blocking
433 	 * memory allocations, and in the file backing case, blocking I/O.
434 	 * This move done routine is generally called in the SIM's
435 	 * interrupt context, and therefore we cannot block.
436 	 */
437 	mtx_lock(&be_lun->queue_lock);
438 	/*
439 	 * XXX KDM make sure that links is okay to use at this point.
440 	 * Otherwise, we either need to add another field to ctl_io_hdr,
441 	 * or deal with resource allocation here.
442 	 */
443 	STAILQ_INSERT_TAIL(&be_lun->datamove_queue, &io->io_hdr, links);
444 	mtx_unlock(&be_lun->queue_lock);
445 
446 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
447 
448 	return (0);
449 }
450 
451 static void
452 ctl_be_block_biodone(struct bio *bio)
453 {
454 	struct ctl_be_block_io *beio;
455 	struct ctl_be_block_lun *be_lun;
456 	union ctl_io *io;
457 	int error;
458 
459 	beio = bio->bio_caller1;
460 	be_lun = beio->lun;
461 	io = beio->io;
462 
463 	DPRINTF("entered\n");
464 
465 	error = bio->bio_error;
466 	mtx_lock(&be_lun->io_lock);
467 	if (error != 0)
468 		beio->num_errors++;
469 
470 	beio->num_bios_done++;
471 
472 	/*
473 	 * XXX KDM will this cause WITNESS to complain?  Holding a lock
474 	 * during the free might cause it to complain.
475 	 */
476 	g_destroy_bio(bio);
477 
478 	/*
479 	 * If the send complete bit isn't set, or we aren't the last I/O to
480 	 * complete, then we're done.
481 	 */
482 	if ((beio->send_complete == 0)
483 	 || (beio->num_bios_done < beio->num_bios_sent)) {
484 		mtx_unlock(&be_lun->io_lock);
485 		return;
486 	}
487 
488 	/*
489 	 * At this point, we've verified that we are the last I/O to
490 	 * complete, so it's safe to drop the lock.
491 	 */
492 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
493 	    beio->ds_tag_type, beio->ds_trans_type,
494 	    /*now*/ NULL, /*then*/&beio->ds_t0);
495 	mtx_unlock(&be_lun->io_lock);
496 
497 	/*
498 	 * If there are any errors from the backing device, we fail the
499 	 * entire I/O with a medium error.
500 	 */
501 	if (beio->num_errors > 0) {
502 		if (error == EOPNOTSUPP) {
503 			ctl_set_invalid_opcode(&io->scsiio);
504 		} else if (beio->bio_cmd == BIO_FLUSH) {
505 			/* XXX KDM is there is a better error here? */
506 			ctl_set_internal_failure(&io->scsiio,
507 						 /*sks_valid*/ 1,
508 						 /*retry_count*/ 0xbad2);
509 		} else
510 			ctl_set_medium_error(&io->scsiio);
511 		ctl_complete_beio(beio);
512 		return;
513 	}
514 
515 	/*
516 	 * If this is a write, a flush, a delete or verify, we're all done.
517 	 * If this is a read, we can now send the data to the user.
518 	 */
519 	if ((beio->bio_cmd == BIO_WRITE)
520 	 || (beio->bio_cmd == BIO_FLUSH)
521 	 || (beio->bio_cmd == BIO_DELETE)
522 	 || (ARGS(io)->flags & CTL_LLF_VERIFY)) {
523 		ctl_set_success(&io->scsiio);
524 		ctl_complete_beio(beio);
525 	} else {
526 #ifdef CTL_TIME_IO
527         	getbintime(&io->io_hdr.dma_start_bt);
528 #endif
529 		ctl_datamove(io);
530 	}
531 }
532 
533 static void
534 ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
535 			struct ctl_be_block_io *beio)
536 {
537 	union ctl_io *io = beio->io;
538 	struct mount *mountpoint;
539 	int error, lock_flags;
540 
541 	DPRINTF("entered\n");
542 
543 	binuptime(&beio->ds_t0);
544 	mtx_lock(&be_lun->io_lock);
545 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
546 	mtx_unlock(&be_lun->io_lock);
547 
548 	(void) vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
549 
550 	if (MNT_SHARED_WRITES(mountpoint)
551 	 || ((mountpoint == NULL)
552 	  && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
553 		lock_flags = LK_SHARED;
554 	else
555 		lock_flags = LK_EXCLUSIVE;
556 
557 	vn_lock(be_lun->vn, lock_flags | LK_RETRY);
558 
559 	error = VOP_FSYNC(be_lun->vn, MNT_WAIT, curthread);
560 	VOP_UNLOCK(be_lun->vn, 0);
561 
562 	vn_finished_write(mountpoint);
563 
564 	mtx_lock(&be_lun->io_lock);
565 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
566 	    beio->ds_tag_type, beio->ds_trans_type,
567 	    /*now*/ NULL, /*then*/&beio->ds_t0);
568 	mtx_unlock(&be_lun->io_lock);
569 
570 	if (error == 0)
571 		ctl_set_success(&io->scsiio);
572 	else {
573 		/* XXX KDM is there is a better error here? */
574 		ctl_set_internal_failure(&io->scsiio,
575 					 /*sks_valid*/ 1,
576 					 /*retry_count*/ 0xbad1);
577 	}
578 
579 	ctl_complete_beio(beio);
580 }
581 
582 SDT_PROBE_DEFINE1(cbb, kernel, read, file_start, "uint64_t");
583 SDT_PROBE_DEFINE1(cbb, kernel, write, file_start, "uint64_t");
584 SDT_PROBE_DEFINE1(cbb, kernel, read, file_done,"uint64_t");
585 SDT_PROBE_DEFINE1(cbb, kernel, write, file_done, "uint64_t");
586 
587 static void
588 ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
589 			   struct ctl_be_block_io *beio)
590 {
591 	struct ctl_be_block_filedata *file_data;
592 	union ctl_io *io;
593 	struct uio xuio;
594 	struct iovec *xiovec;
595 	int flags;
596 	int error, i;
597 
598 	DPRINTF("entered\n");
599 
600 	file_data = &be_lun->backend.file;
601 	io = beio->io;
602 	flags = 0;
603 	if (ARGS(io)->flags & CTL_LLF_DPO)
604 		flags |= IO_DIRECT;
605 	if (beio->bio_cmd == BIO_WRITE && ARGS(io)->flags & CTL_LLF_FUA)
606 		flags |= IO_SYNC;
607 
608 	bzero(&xuio, sizeof(xuio));
609 	if (beio->bio_cmd == BIO_READ) {
610 		SDT_PROBE(cbb, kernel, read, file_start, 0, 0, 0, 0, 0);
611 		xuio.uio_rw = UIO_READ;
612 	} else {
613 		SDT_PROBE(cbb, kernel, write, file_start, 0, 0, 0, 0, 0);
614 		xuio.uio_rw = UIO_WRITE;
615 	}
616 	xuio.uio_offset = beio->io_offset;
617 	xuio.uio_resid = beio->io_len;
618 	xuio.uio_segflg = UIO_SYSSPACE;
619 	xuio.uio_iov = beio->xiovecs;
620 	xuio.uio_iovcnt = beio->num_segs;
621 	xuio.uio_td = curthread;
622 
623 	for (i = 0, xiovec = xuio.uio_iov; i < xuio.uio_iovcnt; i++, xiovec++) {
624 		xiovec->iov_base = beio->sg_segs[i].addr;
625 		xiovec->iov_len = beio->sg_segs[i].len;
626 	}
627 
628 	binuptime(&beio->ds_t0);
629 	mtx_lock(&be_lun->io_lock);
630 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
631 	mtx_unlock(&be_lun->io_lock);
632 
633 	if (beio->bio_cmd == BIO_READ) {
634 		vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
635 
636 		/*
637 		 * UFS pays attention to IO_DIRECT for reads.  If the
638 		 * DIRECTIO option is configured into the kernel, it calls
639 		 * ffs_rawread().  But that only works for single-segment
640 		 * uios with user space addresses.  In our case, with a
641 		 * kernel uio, it still reads into the buffer cache, but it
642 		 * will just try to release the buffer from the cache later
643 		 * on in ffs_read().
644 		 *
645 		 * ZFS does not pay attention to IO_DIRECT for reads.
646 		 *
647 		 * UFS does not pay attention to IO_SYNC for reads.
648 		 *
649 		 * ZFS pays attention to IO_SYNC (which translates into the
650 		 * Solaris define FRSYNC for zfs_read()) for reads.  It
651 		 * attempts to sync the file before reading.
652 		 *
653 		 * So, to attempt to provide some barrier semantics in the
654 		 * BIO_ORDERED case, set both IO_DIRECT and IO_SYNC.
655 		 */
656 		error = VOP_READ(be_lun->vn, &xuio, flags, file_data->cred);
657 
658 		VOP_UNLOCK(be_lun->vn, 0);
659 		SDT_PROBE(cbb, kernel, read, file_done, 0, 0, 0, 0, 0);
660 	} else {
661 		struct mount *mountpoint;
662 		int lock_flags;
663 
664 		(void)vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
665 
666 		if (MNT_SHARED_WRITES(mountpoint)
667 		 || ((mountpoint == NULL)
668 		  && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
669 			lock_flags = LK_SHARED;
670 		else
671 			lock_flags = LK_EXCLUSIVE;
672 
673 		vn_lock(be_lun->vn, lock_flags | LK_RETRY);
674 
675 		/*
676 		 * UFS pays attention to IO_DIRECT for writes.  The write
677 		 * is done asynchronously.  (Normally the write would just
678 		 * get put into cache.
679 		 *
680 		 * UFS pays attention to IO_SYNC for writes.  It will
681 		 * attempt to write the buffer out synchronously if that
682 		 * flag is set.
683 		 *
684 		 * ZFS does not pay attention to IO_DIRECT for writes.
685 		 *
686 		 * ZFS pays attention to IO_SYNC (a.k.a. FSYNC or FRSYNC)
687 		 * for writes.  It will flush the transaction from the
688 		 * cache before returning.
689 		 *
690 		 * So if we've got the BIO_ORDERED flag set, we want
691 		 * IO_SYNC in either the UFS or ZFS case.
692 		 */
693 		error = VOP_WRITE(be_lun->vn, &xuio, flags, file_data->cred);
694 		VOP_UNLOCK(be_lun->vn, 0);
695 
696 		vn_finished_write(mountpoint);
697 		SDT_PROBE(cbb, kernel, write, file_done, 0, 0, 0, 0, 0);
698         }
699 
700 	mtx_lock(&be_lun->io_lock);
701 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
702 	    beio->ds_tag_type, beio->ds_trans_type,
703 	    /*now*/ NULL, /*then*/&beio->ds_t0);
704 	mtx_unlock(&be_lun->io_lock);
705 
706 	/*
707 	 * If we got an error, set the sense data to "MEDIUM ERROR" and
708 	 * return the I/O to the user.
709 	 */
710 	if (error != 0) {
711 		char path_str[32];
712 
713 		ctl_scsi_path_string(io, path_str, sizeof(path_str));
714 		/*
715 		 * XXX KDM ZFS returns ENOSPC when the underlying
716 		 * filesystem fills up.  What kind of SCSI error should we
717 		 * return for that?
718 		 */
719 		printf("%s%s command returned errno %d\n", path_str,
720 		       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE", error);
721 		ctl_set_medium_error(&io->scsiio);
722 		ctl_complete_beio(beio);
723 		return;
724 	}
725 
726 	/*
727 	 * If this is a write or a verify, we're all done.
728 	 * If this is a read, we can now send the data to the user.
729 	 */
730 	if ((beio->bio_cmd == BIO_WRITE) ||
731 	    (ARGS(io)->flags & CTL_LLF_VERIFY)) {
732 		ctl_set_success(&io->scsiio);
733 		ctl_complete_beio(beio);
734 	} else {
735 #ifdef CTL_TIME_IO
736         	getbintime(&io->io_hdr.dma_start_bt);
737 #endif
738 		ctl_datamove(io);
739 	}
740 }
741 
742 static void
743 ctl_be_block_dispatch_zvol(struct ctl_be_block_lun *be_lun,
744 			   struct ctl_be_block_io *beio)
745 {
746 	struct ctl_be_block_devdata *dev_data;
747 	union ctl_io *io;
748 	struct uio xuio;
749 	struct iovec *xiovec;
750 	int flags;
751 	int error, i;
752 
753 	DPRINTF("entered\n");
754 
755 	dev_data = &be_lun->backend.dev;
756 	io = beio->io;
757 	flags = 0;
758 	if (ARGS(io)->flags & CTL_LLF_DPO)
759 		flags |= IO_DIRECT;
760 	if (beio->bio_cmd == BIO_WRITE && ARGS(io)->flags & CTL_LLF_FUA)
761 		flags |= IO_SYNC;
762 
763 	bzero(&xuio, sizeof(xuio));
764 	if (beio->bio_cmd == BIO_READ) {
765 		SDT_PROBE(cbb, kernel, read, file_start, 0, 0, 0, 0, 0);
766 		xuio.uio_rw = UIO_READ;
767 	} else {
768 		SDT_PROBE(cbb, kernel, write, file_start, 0, 0, 0, 0, 0);
769 		xuio.uio_rw = UIO_WRITE;
770 	}
771 	xuio.uio_offset = beio->io_offset;
772 	xuio.uio_resid = beio->io_len;
773 	xuio.uio_segflg = UIO_SYSSPACE;
774 	xuio.uio_iov = beio->xiovecs;
775 	xuio.uio_iovcnt = beio->num_segs;
776 	xuio.uio_td = curthread;
777 
778 	for (i = 0, xiovec = xuio.uio_iov; i < xuio.uio_iovcnt; i++, xiovec++) {
779 		xiovec->iov_base = beio->sg_segs[i].addr;
780 		xiovec->iov_len = beio->sg_segs[i].len;
781 	}
782 
783 	binuptime(&beio->ds_t0);
784 	mtx_lock(&be_lun->io_lock);
785 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
786 	mtx_unlock(&be_lun->io_lock);
787 
788 	if (beio->bio_cmd == BIO_READ) {
789 		error = (*dev_data->csw->d_read)(dev_data->cdev, &xuio, flags);
790 		SDT_PROBE(cbb, kernel, read, file_done, 0, 0, 0, 0, 0);
791 	} else {
792 		error = (*dev_data->csw->d_write)(dev_data->cdev, &xuio, flags);
793 		SDT_PROBE(cbb, kernel, write, file_done, 0, 0, 0, 0, 0);
794 	}
795 
796 	mtx_lock(&be_lun->io_lock);
797 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
798 	    beio->ds_tag_type, beio->ds_trans_type,
799 	    /*now*/ NULL, /*then*/&beio->ds_t0);
800 	mtx_unlock(&be_lun->io_lock);
801 
802 	/*
803 	 * If we got an error, set the sense data to "MEDIUM ERROR" and
804 	 * return the I/O to the user.
805 	 */
806 	if (error != 0) {
807 		ctl_set_medium_error(&io->scsiio);
808 		ctl_complete_beio(beio);
809 		return;
810 	}
811 
812 	/*
813 	 * If this is a write or a verify, we're all done.
814 	 * If this is a read, we can now send the data to the user.
815 	 */
816 	if ((beio->bio_cmd == BIO_WRITE) ||
817 	    (ARGS(io)->flags & CTL_LLF_VERIFY)) {
818 		ctl_set_success(&io->scsiio);
819 		ctl_complete_beio(beio);
820 	} else {
821 #ifdef CTL_TIME_IO
822         	getbintime(&io->io_hdr.dma_start_bt);
823 #endif
824 		ctl_datamove(io);
825 	}
826 }
827 
828 static void
829 ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
830 		       struct ctl_be_block_io *beio)
831 {
832 	struct bio *bio;
833 	union ctl_io *io;
834 	struct ctl_be_block_devdata *dev_data;
835 
836 	dev_data = &be_lun->backend.dev;
837 	io = beio->io;
838 
839 	DPRINTF("entered\n");
840 
841 	/* This can't fail, it's a blocking allocation. */
842 	bio = g_alloc_bio();
843 
844 	bio->bio_cmd	    = BIO_FLUSH;
845 	bio->bio_flags	   |= BIO_ORDERED;
846 	bio->bio_dev	    = dev_data->cdev;
847 	bio->bio_offset	    = 0;
848 	bio->bio_data	    = 0;
849 	bio->bio_done	    = ctl_be_block_biodone;
850 	bio->bio_caller1    = beio;
851 	bio->bio_pblkno	    = 0;
852 
853 	/*
854 	 * We don't need to acquire the LUN lock here, because we are only
855 	 * sending one bio, and so there is no other context to synchronize
856 	 * with.
857 	 */
858 	beio->num_bios_sent = 1;
859 	beio->send_complete = 1;
860 
861 	binuptime(&beio->ds_t0);
862 	mtx_lock(&be_lun->io_lock);
863 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
864 	mtx_unlock(&be_lun->io_lock);
865 
866 	(*dev_data->csw->d_strategy)(bio);
867 }
868 
869 static void
870 ctl_be_block_unmap_dev_range(struct ctl_be_block_lun *be_lun,
871 		       struct ctl_be_block_io *beio,
872 		       uint64_t off, uint64_t len, int last)
873 {
874 	struct bio *bio;
875 	struct ctl_be_block_devdata *dev_data;
876 	uint64_t maxlen;
877 
878 	dev_data = &be_lun->backend.dev;
879 	maxlen = LONG_MAX - (LONG_MAX % be_lun->blocksize);
880 	while (len > 0) {
881 		bio = g_alloc_bio();
882 		bio->bio_cmd	    = BIO_DELETE;
883 		bio->bio_dev	    = dev_data->cdev;
884 		bio->bio_offset	    = off;
885 		bio->bio_length	    = MIN(len, maxlen);
886 		bio->bio_data	    = 0;
887 		bio->bio_done	    = ctl_be_block_biodone;
888 		bio->bio_caller1    = beio;
889 		bio->bio_pblkno     = off / be_lun->blocksize;
890 
891 		off += bio->bio_length;
892 		len -= bio->bio_length;
893 
894 		mtx_lock(&be_lun->io_lock);
895 		beio->num_bios_sent++;
896 		if (last && len == 0)
897 			beio->send_complete = 1;
898 		mtx_unlock(&be_lun->io_lock);
899 
900 		(*dev_data->csw->d_strategy)(bio);
901 	}
902 }
903 
904 static void
905 ctl_be_block_unmap_dev(struct ctl_be_block_lun *be_lun,
906 		       struct ctl_be_block_io *beio)
907 {
908 	union ctl_io *io;
909 	struct ctl_be_block_devdata *dev_data;
910 	struct ctl_ptr_len_flags *ptrlen;
911 	struct scsi_unmap_desc *buf, *end;
912 	uint64_t len;
913 
914 	dev_data = &be_lun->backend.dev;
915 	io = beio->io;
916 
917 	DPRINTF("entered\n");
918 
919 	binuptime(&beio->ds_t0);
920 	mtx_lock(&be_lun->io_lock);
921 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
922 	mtx_unlock(&be_lun->io_lock);
923 
924 	if (beio->io_offset == -1) {
925 		beio->io_len = 0;
926 		ptrlen = (struct ctl_ptr_len_flags *)&io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN];
927 		buf = (struct scsi_unmap_desc *)ptrlen->ptr;
928 		end = buf + ptrlen->len / sizeof(*buf);
929 		for (; buf < end; buf++) {
930 			len = (uint64_t)scsi_4btoul(buf->length) *
931 			    be_lun->blocksize;
932 			beio->io_len += len;
933 			ctl_be_block_unmap_dev_range(be_lun, beio,
934 			    scsi_8btou64(buf->lba) * be_lun->blocksize, len,
935 			    (end - buf < 2) ? TRUE : FALSE);
936 		}
937 	} else
938 		ctl_be_block_unmap_dev_range(be_lun, beio,
939 		    beio->io_offset, beio->io_len, TRUE);
940 }
941 
942 static void
943 ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
944 			  struct ctl_be_block_io *beio)
945 {
946 	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
947 	int i;
948 	struct bio *bio;
949 	struct ctl_be_block_devdata *dev_data;
950 	off_t cur_offset;
951 	int max_iosize;
952 
953 	DPRINTF("entered\n");
954 
955 	dev_data = &be_lun->backend.dev;
956 
957 	/*
958 	 * We have to limit our I/O size to the maximum supported by the
959 	 * backend device.  Hopefully it is MAXPHYS.  If the driver doesn't
960 	 * set it properly, use DFLTPHYS.
961 	 */
962 	max_iosize = dev_data->cdev->si_iosize_max;
963 	if (max_iosize < PAGE_SIZE)
964 		max_iosize = DFLTPHYS;
965 
966 	cur_offset = beio->io_offset;
967 	for (i = 0; i < beio->num_segs; i++) {
968 		size_t cur_size;
969 		uint8_t *cur_ptr;
970 
971 		cur_size = beio->sg_segs[i].len;
972 		cur_ptr = beio->sg_segs[i].addr;
973 
974 		while (cur_size > 0) {
975 			/* This can't fail, it's a blocking allocation. */
976 			bio = g_alloc_bio();
977 
978 			KASSERT(bio != NULL, ("g_alloc_bio() failed!\n"));
979 
980 			bio->bio_cmd = beio->bio_cmd;
981 			bio->bio_dev = dev_data->cdev;
982 			bio->bio_caller1 = beio;
983 			bio->bio_length = min(cur_size, max_iosize);
984 			bio->bio_offset = cur_offset;
985 			bio->bio_data = cur_ptr;
986 			bio->bio_done = ctl_be_block_biodone;
987 			bio->bio_pblkno = cur_offset / be_lun->blocksize;
988 
989 			cur_offset += bio->bio_length;
990 			cur_ptr += bio->bio_length;
991 			cur_size -= bio->bio_length;
992 
993 			TAILQ_INSERT_TAIL(&queue, bio, bio_queue);
994 			beio->num_bios_sent++;
995 		}
996 	}
997 	binuptime(&beio->ds_t0);
998 	mtx_lock(&be_lun->io_lock);
999 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
1000 	beio->send_complete = 1;
1001 	mtx_unlock(&be_lun->io_lock);
1002 
1003 	/*
1004 	 * Fire off all allocated requests!
1005 	 */
1006 	while ((bio = TAILQ_FIRST(&queue)) != NULL) {
1007 		TAILQ_REMOVE(&queue, bio, bio_queue);
1008 		(*dev_data->csw->d_strategy)(bio);
1009 	}
1010 }
1011 
1012 static void
1013 ctl_be_block_cw_done_ws(struct ctl_be_block_io *beio)
1014 {
1015 	union ctl_io *io;
1016 
1017 	io = beio->io;
1018 	ctl_free_beio(beio);
1019 	if ((io->io_hdr.flags & CTL_FLAG_ABORT) ||
1020 	    ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE &&
1021 	     (io->io_hdr.status & CTL_STATUS_MASK) != CTL_SUCCESS)) {
1022 		ctl_config_write_done(io);
1023 		return;
1024 	}
1025 
1026 	ctl_be_block_config_write(io);
1027 }
1028 
1029 static void
1030 ctl_be_block_cw_dispatch_ws(struct ctl_be_block_lun *be_lun,
1031 			    union ctl_io *io)
1032 {
1033 	struct ctl_be_block_io *beio;
1034 	struct ctl_be_block_softc *softc;
1035 	struct ctl_lba_len_flags *lbalen;
1036 	uint64_t len_left, lba;
1037 	int i, seglen;
1038 	uint8_t *buf, *end;
1039 
1040 	DPRINTF("entered\n");
1041 
1042 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1043 	softc = be_lun->softc;
1044 	lbalen = ARGS(beio->io);
1045 
1046 	if (lbalen->flags & ~(SWS_LBDATA | SWS_UNMAP | SWS_ANCHOR | SWS_NDOB) ||
1047 	    (lbalen->flags & (SWS_UNMAP | SWS_ANCHOR) && be_lun->unmap == NULL)) {
1048 		ctl_free_beio(beio);
1049 		ctl_set_invalid_field(&io->scsiio,
1050 				      /*sks_valid*/ 1,
1051 				      /*command*/ 1,
1052 				      /*field*/ 1,
1053 				      /*bit_valid*/ 0,
1054 				      /*bit*/ 0);
1055 		ctl_config_write_done(io);
1056 		return;
1057 	}
1058 
1059 	switch (io->scsiio.tag_type) {
1060 	case CTL_TAG_ORDERED:
1061 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1062 		break;
1063 	case CTL_TAG_HEAD_OF_QUEUE:
1064 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1065 		break;
1066 	case CTL_TAG_UNTAGGED:
1067 	case CTL_TAG_SIMPLE:
1068 	case CTL_TAG_ACA:
1069 	default:
1070 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1071 		break;
1072 	}
1073 
1074 	if (lbalen->flags & (SWS_UNMAP | SWS_ANCHOR)) {
1075 		beio->io_offset = lbalen->lba * be_lun->blocksize;
1076 		beio->io_len = (uint64_t)lbalen->len * be_lun->blocksize;
1077 		beio->bio_cmd = BIO_DELETE;
1078 		beio->ds_trans_type = DEVSTAT_FREE;
1079 
1080 		be_lun->unmap(be_lun, beio);
1081 		return;
1082 	}
1083 
1084 	beio->bio_cmd = BIO_WRITE;
1085 	beio->ds_trans_type = DEVSTAT_WRITE;
1086 
1087 	DPRINTF("WRITE SAME at LBA %jx len %u\n",
1088 	       (uintmax_t)lbalen->lba, lbalen->len);
1089 
1090 	len_left = (uint64_t)lbalen->len * be_lun->blocksize;
1091 	for (i = 0, lba = 0; i < CTLBLK_MAX_SEGS && len_left > 0; i++) {
1092 
1093 		/*
1094 		 * Setup the S/G entry for this chunk.
1095 		 */
1096 		seglen = MIN(CTLBLK_MAX_SEG, len_left);
1097 		seglen -= seglen % be_lun->blocksize;
1098 		beio->sg_segs[i].len = seglen;
1099 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1100 
1101 		DPRINTF("segment %d addr %p len %zd\n", i,
1102 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1103 
1104 		beio->num_segs++;
1105 		len_left -= seglen;
1106 
1107 		buf = beio->sg_segs[i].addr;
1108 		end = buf + seglen;
1109 		for (; buf < end; buf += be_lun->blocksize) {
1110 			memcpy(buf, io->scsiio.kern_data_ptr, be_lun->blocksize);
1111 			if (lbalen->flags & SWS_LBDATA)
1112 				scsi_ulto4b(lbalen->lba + lba, buf);
1113 			lba++;
1114 		}
1115 	}
1116 
1117 	beio->io_offset = lbalen->lba * be_lun->blocksize;
1118 	beio->io_len = lba * be_lun->blocksize;
1119 
1120 	/* We can not do all in one run. Correct and schedule rerun. */
1121 	if (len_left > 0) {
1122 		lbalen->lba += lba;
1123 		lbalen->len -= lba;
1124 		beio->beio_cont = ctl_be_block_cw_done_ws;
1125 	}
1126 
1127 	be_lun->dispatch(be_lun, beio);
1128 }
1129 
1130 static void
1131 ctl_be_block_cw_dispatch_unmap(struct ctl_be_block_lun *be_lun,
1132 			    union ctl_io *io)
1133 {
1134 	struct ctl_be_block_io *beio;
1135 	struct ctl_be_block_softc *softc;
1136 	struct ctl_ptr_len_flags *ptrlen;
1137 
1138 	DPRINTF("entered\n");
1139 
1140 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1141 	softc = be_lun->softc;
1142 	ptrlen = (struct ctl_ptr_len_flags *)&io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN];
1143 
1144 	if ((ptrlen->flags & ~SU_ANCHOR) != 0 || be_lun->unmap == NULL) {
1145 		ctl_free_beio(beio);
1146 		ctl_set_invalid_field(&io->scsiio,
1147 				      /*sks_valid*/ 0,
1148 				      /*command*/ 1,
1149 				      /*field*/ 0,
1150 				      /*bit_valid*/ 0,
1151 				      /*bit*/ 0);
1152 		ctl_config_write_done(io);
1153 		return;
1154 	}
1155 
1156 	switch (io->scsiio.tag_type) {
1157 	case CTL_TAG_ORDERED:
1158 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1159 		break;
1160 	case CTL_TAG_HEAD_OF_QUEUE:
1161 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1162 		break;
1163 	case CTL_TAG_UNTAGGED:
1164 	case CTL_TAG_SIMPLE:
1165 	case CTL_TAG_ACA:
1166 	default:
1167 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1168 		break;
1169 	}
1170 
1171 	beio->io_len = 0;
1172 	beio->io_offset = -1;
1173 
1174 	beio->bio_cmd = BIO_DELETE;
1175 	beio->ds_trans_type = DEVSTAT_FREE;
1176 
1177 	DPRINTF("UNMAP\n");
1178 
1179 	be_lun->unmap(be_lun, beio);
1180 }
1181 
1182 static void
1183 ctl_be_block_cw_done(struct ctl_be_block_io *beio)
1184 {
1185 	union ctl_io *io;
1186 
1187 	io = beio->io;
1188 	ctl_free_beio(beio);
1189 	ctl_config_write_done(io);
1190 }
1191 
1192 static void
1193 ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
1194 			 union ctl_io *io)
1195 {
1196 	struct ctl_be_block_io *beio;
1197 	struct ctl_be_block_softc *softc;
1198 
1199 	DPRINTF("entered\n");
1200 
1201 	softc = be_lun->softc;
1202 	beio = ctl_alloc_beio(softc);
1203 	beio->io = io;
1204 	beio->lun = be_lun;
1205 	beio->beio_cont = ctl_be_block_cw_done;
1206 	PRIV(io)->ptr = (void *)beio;
1207 
1208 	switch (io->scsiio.cdb[0]) {
1209 	case SYNCHRONIZE_CACHE:
1210 	case SYNCHRONIZE_CACHE_16:
1211 		beio->bio_cmd = BIO_FLUSH;
1212 		beio->ds_trans_type = DEVSTAT_NO_DATA;
1213 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1214 		beio->io_len = 0;
1215 		be_lun->lun_flush(be_lun, beio);
1216 		break;
1217 	case WRITE_SAME_10:
1218 	case WRITE_SAME_16:
1219 		ctl_be_block_cw_dispatch_ws(be_lun, io);
1220 		break;
1221 	case UNMAP:
1222 		ctl_be_block_cw_dispatch_unmap(be_lun, io);
1223 		break;
1224 	default:
1225 		panic("Unhandled CDB type %#x", io->scsiio.cdb[0]);
1226 		break;
1227 	}
1228 }
1229 
1230 SDT_PROBE_DEFINE1(cbb, kernel, read, start, "uint64_t");
1231 SDT_PROBE_DEFINE1(cbb, kernel, write, start, "uint64_t");
1232 SDT_PROBE_DEFINE1(cbb, kernel, read, alloc_done, "uint64_t");
1233 SDT_PROBE_DEFINE1(cbb, kernel, write, alloc_done, "uint64_t");
1234 
1235 static void
1236 ctl_be_block_next(struct ctl_be_block_io *beio)
1237 {
1238 	struct ctl_be_block_lun *be_lun;
1239 	union ctl_io *io;
1240 
1241 	io = beio->io;
1242 	be_lun = beio->lun;
1243 	ctl_free_beio(beio);
1244 	if ((io->io_hdr.flags & CTL_FLAG_ABORT) ||
1245 	    ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE &&
1246 	     (io->io_hdr.status & CTL_STATUS_MASK) != CTL_SUCCESS)) {
1247 		ctl_data_submit_done(io);
1248 		return;
1249 	}
1250 
1251 	io->io_hdr.status &= ~CTL_STATUS_MASK;
1252 	io->io_hdr.status |= CTL_STATUS_NONE;
1253 
1254 	mtx_lock(&be_lun->queue_lock);
1255 	/*
1256 	 * XXX KDM make sure that links is okay to use at this point.
1257 	 * Otherwise, we either need to add another field to ctl_io_hdr,
1258 	 * or deal with resource allocation here.
1259 	 */
1260 	STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1261 	mtx_unlock(&be_lun->queue_lock);
1262 
1263 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1264 }
1265 
1266 static void
1267 ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
1268 			   union ctl_io *io)
1269 {
1270 	struct ctl_be_block_io *beio;
1271 	struct ctl_be_block_softc *softc;
1272 	struct ctl_lba_len_flags *lbalen;
1273 	struct ctl_ptr_len_flags *bptrlen;
1274 	uint64_t len_left, lbas;
1275 	int i;
1276 
1277 	softc = be_lun->softc;
1278 
1279 	DPRINTF("entered\n");
1280 
1281 	lbalen = ARGS(io);
1282 	if (lbalen->flags & CTL_LLF_WRITE) {
1283 		SDT_PROBE(cbb, kernel, write, start, 0, 0, 0, 0, 0);
1284 	} else {
1285 		SDT_PROBE(cbb, kernel, read, start, 0, 0, 0, 0, 0);
1286 	}
1287 
1288 	beio = ctl_alloc_beio(softc);
1289 	beio->io = io;
1290 	beio->lun = be_lun;
1291 	bptrlen = PRIV(io);
1292 	bptrlen->ptr = (void *)beio;
1293 
1294 	switch (io->scsiio.tag_type) {
1295 	case CTL_TAG_ORDERED:
1296 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1297 		break;
1298 	case CTL_TAG_HEAD_OF_QUEUE:
1299 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1300 		break;
1301 	case CTL_TAG_UNTAGGED:
1302 	case CTL_TAG_SIMPLE:
1303 	case CTL_TAG_ACA:
1304 	default:
1305 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1306 		break;
1307 	}
1308 
1309 	if (lbalen->flags & CTL_LLF_WRITE) {
1310 		beio->bio_cmd = BIO_WRITE;
1311 		beio->ds_trans_type = DEVSTAT_WRITE;
1312 	} else {
1313 		beio->bio_cmd = BIO_READ;
1314 		beio->ds_trans_type = DEVSTAT_READ;
1315 	}
1316 
1317 	DPRINTF("%s at LBA %jx len %u @%ju\n",
1318 	       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE",
1319 	       (uintmax_t)lbalen->lba, lbalen->len, bptrlen->len);
1320 	if (lbalen->flags & CTL_LLF_COMPARE)
1321 		lbas = CTLBLK_HALF_IO_SIZE;
1322 	else
1323 		lbas = CTLBLK_MAX_IO_SIZE;
1324 	lbas = MIN(lbalen->len - bptrlen->len, lbas / be_lun->blocksize);
1325 	beio->io_offset = (lbalen->lba + bptrlen->len) * be_lun->blocksize;
1326 	beio->io_len = lbas * be_lun->blocksize;
1327 	bptrlen->len += lbas;
1328 
1329 	for (i = 0, len_left = beio->io_len; len_left > 0; i++) {
1330 		KASSERT(i < CTLBLK_MAX_SEGS, ("Too many segs (%d >= %d)",
1331 		    i, CTLBLK_MAX_SEGS));
1332 
1333 		/*
1334 		 * Setup the S/G entry for this chunk.
1335 		 */
1336 		beio->sg_segs[i].len = min(CTLBLK_MAX_SEG, len_left);
1337 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1338 
1339 		DPRINTF("segment %d addr %p len %zd\n", i,
1340 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1341 
1342 		/* Set up second segment for compare operation. */
1343 		if (lbalen->flags & CTL_LLF_COMPARE) {
1344 			beio->sg_segs[i + CTLBLK_HALF_SEGS].len =
1345 			    beio->sg_segs[i].len;
1346 			beio->sg_segs[i + CTLBLK_HALF_SEGS].addr =
1347 			    uma_zalloc(be_lun->lun_zone, M_WAITOK);
1348 		}
1349 
1350 		beio->num_segs++;
1351 		len_left -= beio->sg_segs[i].len;
1352 	}
1353 	if (bptrlen->len < lbalen->len)
1354 		beio->beio_cont = ctl_be_block_next;
1355 	io->scsiio.be_move_done = ctl_be_block_move_done;
1356 	/* For compare we have separate S/G lists for read and datamove. */
1357 	if (lbalen->flags & CTL_LLF_COMPARE)
1358 		io->scsiio.kern_data_ptr = (uint8_t *)&beio->sg_segs[CTLBLK_HALF_SEGS];
1359 	else
1360 		io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
1361 	io->scsiio.kern_data_len = beio->io_len;
1362 	io->scsiio.kern_data_resid = 0;
1363 	io->scsiio.kern_sg_entries = beio->num_segs;
1364 	io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
1365 
1366 	/*
1367 	 * For the read case, we need to read the data into our buffers and
1368 	 * then we can send it back to the user.  For the write case, we
1369 	 * need to get the data from the user first.
1370 	 */
1371 	if (beio->bio_cmd == BIO_READ) {
1372 		SDT_PROBE(cbb, kernel, read, alloc_done, 0, 0, 0, 0, 0);
1373 		be_lun->dispatch(be_lun, beio);
1374 	} else {
1375 		SDT_PROBE(cbb, kernel, write, alloc_done, 0, 0, 0, 0, 0);
1376 #ifdef CTL_TIME_IO
1377         	getbintime(&io->io_hdr.dma_start_bt);
1378 #endif
1379 		ctl_datamove(io);
1380 	}
1381 }
1382 
1383 static void
1384 ctl_be_block_worker(void *context, int pending)
1385 {
1386 	struct ctl_be_block_lun *be_lun;
1387 	struct ctl_be_block_softc *softc;
1388 	union ctl_io *io;
1389 
1390 	be_lun = (struct ctl_be_block_lun *)context;
1391 	softc = be_lun->softc;
1392 
1393 	DPRINTF("entered\n");
1394 
1395 	mtx_lock(&be_lun->queue_lock);
1396 	for (;;) {
1397 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->datamove_queue);
1398 		if (io != NULL) {
1399 			struct ctl_be_block_io *beio;
1400 
1401 			DPRINTF("datamove queue\n");
1402 
1403 			STAILQ_REMOVE(&be_lun->datamove_queue, &io->io_hdr,
1404 				      ctl_io_hdr, links);
1405 
1406 			mtx_unlock(&be_lun->queue_lock);
1407 
1408 			beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1409 
1410 			be_lun->dispatch(be_lun, beio);
1411 
1412 			mtx_lock(&be_lun->queue_lock);
1413 			continue;
1414 		}
1415 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->config_write_queue);
1416 		if (io != NULL) {
1417 
1418 			DPRINTF("config write queue\n");
1419 
1420 			STAILQ_REMOVE(&be_lun->config_write_queue, &io->io_hdr,
1421 				      ctl_io_hdr, links);
1422 
1423 			mtx_unlock(&be_lun->queue_lock);
1424 
1425 			ctl_be_block_cw_dispatch(be_lun, io);
1426 
1427 			mtx_lock(&be_lun->queue_lock);
1428 			continue;
1429 		}
1430 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->input_queue);
1431 		if (io != NULL) {
1432 			DPRINTF("input queue\n");
1433 
1434 			STAILQ_REMOVE(&be_lun->input_queue, &io->io_hdr,
1435 				      ctl_io_hdr, links);
1436 			mtx_unlock(&be_lun->queue_lock);
1437 
1438 			/*
1439 			 * We must drop the lock, since this routine and
1440 			 * its children may sleep.
1441 			 */
1442 			ctl_be_block_dispatch(be_lun, io);
1443 
1444 			mtx_lock(&be_lun->queue_lock);
1445 			continue;
1446 		}
1447 
1448 		/*
1449 		 * If we get here, there is no work left in the queues, so
1450 		 * just break out and let the task queue go to sleep.
1451 		 */
1452 		break;
1453 	}
1454 	mtx_unlock(&be_lun->queue_lock);
1455 }
1456 
1457 /*
1458  * Entry point from CTL to the backend for I/O.  We queue everything to a
1459  * work thread, so this just puts the I/O on a queue and wakes up the
1460  * thread.
1461  */
1462 static int
1463 ctl_be_block_submit(union ctl_io *io)
1464 {
1465 	struct ctl_be_block_lun *be_lun;
1466 	struct ctl_be_lun *ctl_be_lun;
1467 
1468 	DPRINTF("entered\n");
1469 
1470 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
1471 		CTL_PRIV_BACKEND_LUN].ptr;
1472 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
1473 
1474 	/*
1475 	 * Make sure we only get SCSI I/O.
1476 	 */
1477 	KASSERT(io->io_hdr.io_type == CTL_IO_SCSI, ("Non-SCSI I/O (type "
1478 		"%#x) encountered", io->io_hdr.io_type));
1479 
1480 	PRIV(io)->len = 0;
1481 
1482 	mtx_lock(&be_lun->queue_lock);
1483 	/*
1484 	 * XXX KDM make sure that links is okay to use at this point.
1485 	 * Otherwise, we either need to add another field to ctl_io_hdr,
1486 	 * or deal with resource allocation here.
1487 	 */
1488 	STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1489 	mtx_unlock(&be_lun->queue_lock);
1490 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1491 
1492 	return (CTL_RETVAL_COMPLETE);
1493 }
1494 
1495 static int
1496 ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
1497 			int flag, struct thread *td)
1498 {
1499 	struct ctl_be_block_softc *softc;
1500 	int error;
1501 
1502 	softc = &backend_block_softc;
1503 
1504 	error = 0;
1505 
1506 	switch (cmd) {
1507 	case CTL_LUN_REQ: {
1508 		struct ctl_lun_req *lun_req;
1509 
1510 		lun_req = (struct ctl_lun_req *)addr;
1511 
1512 		switch (lun_req->reqtype) {
1513 		case CTL_LUNREQ_CREATE:
1514 			error = ctl_be_block_create(softc, lun_req);
1515 			break;
1516 		case CTL_LUNREQ_RM:
1517 			error = ctl_be_block_rm(softc, lun_req);
1518 			break;
1519 		case CTL_LUNREQ_MODIFY:
1520 			error = ctl_be_block_modify(softc, lun_req);
1521 			break;
1522 		default:
1523 			lun_req->status = CTL_LUN_ERROR;
1524 			snprintf(lun_req->error_str, sizeof(lun_req->error_str),
1525 				 "invalid LUN request type %d",
1526 				 lun_req->reqtype);
1527 			break;
1528 		}
1529 		break;
1530 	}
1531 	default:
1532 		error = ENOTTY;
1533 		break;
1534 	}
1535 
1536 	return (error);
1537 }
1538 
1539 static int
1540 ctl_be_block_open_file(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1541 {
1542 	struct ctl_be_block_filedata *file_data;
1543 	struct ctl_lun_create_params *params;
1544 	struct vattr		      vattr;
1545 	int			      error;
1546 
1547 	error = 0;
1548 	file_data = &be_lun->backend.file;
1549 	params = &be_lun->params;
1550 
1551 	be_lun->dev_type = CTL_BE_BLOCK_FILE;
1552 	be_lun->dispatch = ctl_be_block_dispatch_file;
1553 	be_lun->lun_flush = ctl_be_block_flush_file;
1554 
1555 	error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
1556 	if (error != 0) {
1557 		snprintf(req->error_str, sizeof(req->error_str),
1558 			 "error calling VOP_GETATTR() for file %s",
1559 			 be_lun->dev_path);
1560 		return (error);
1561 	}
1562 
1563 	/*
1564 	 * Verify that we have the ability to upgrade to exclusive
1565 	 * access on this file so we can trap errors at open instead
1566 	 * of reporting them during first access.
1567 	 */
1568 	if (VOP_ISLOCKED(be_lun->vn) != LK_EXCLUSIVE) {
1569 		vn_lock(be_lun->vn, LK_UPGRADE | LK_RETRY);
1570 		if (be_lun->vn->v_iflag & VI_DOOMED) {
1571 			error = EBADF;
1572 			snprintf(req->error_str, sizeof(req->error_str),
1573 				 "error locking file %s", be_lun->dev_path);
1574 			return (error);
1575 		}
1576 	}
1577 
1578 
1579 	file_data->cred = crhold(curthread->td_ucred);
1580 	if (params->lun_size_bytes != 0)
1581 		be_lun->size_bytes = params->lun_size_bytes;
1582 	else
1583 		be_lun->size_bytes = vattr.va_size;
1584 	/*
1585 	 * We set the multi thread flag for file operations because all
1586 	 * filesystems (in theory) are capable of allowing multiple readers
1587 	 * of a file at once.  So we want to get the maximum possible
1588 	 * concurrency.
1589 	 */
1590 	be_lun->flags |= CTL_BE_BLOCK_LUN_MULTI_THREAD;
1591 
1592 	/*
1593 	 * XXX KDM vattr.va_blocksize may be larger than 512 bytes here.
1594 	 * With ZFS, it is 131072 bytes.  Block sizes that large don't work
1595 	 * with disklabel and UFS on FreeBSD at least.  Large block sizes
1596 	 * may not work with other OSes as well.  So just export a sector
1597 	 * size of 512 bytes, which should work with any OS or
1598 	 * application.  Since our backing is a file, any block size will
1599 	 * work fine for the backing store.
1600 	 */
1601 #if 0
1602 	be_lun->blocksize= vattr.va_blocksize;
1603 #endif
1604 	if (params->blocksize_bytes != 0)
1605 		be_lun->blocksize = params->blocksize_bytes;
1606 	else
1607 		be_lun->blocksize = 512;
1608 
1609 	/*
1610 	 * Sanity check.  The media size has to be at least one
1611 	 * sector long.
1612 	 */
1613 	if (be_lun->size_bytes < be_lun->blocksize) {
1614 		error = EINVAL;
1615 		snprintf(req->error_str, sizeof(req->error_str),
1616 			 "file %s size %ju < block size %u", be_lun->dev_path,
1617 			 (uintmax_t)be_lun->size_bytes, be_lun->blocksize);
1618 	}
1619 	return (error);
1620 }
1621 
1622 static int
1623 ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1624 {
1625 	struct ctl_lun_create_params *params;
1626 	struct vattr		      vattr;
1627 	struct cdev		     *dev;
1628 	struct cdevsw		     *devsw;
1629 	int			      error;
1630 	off_t			      ps, pss, po, pos;
1631 
1632 	params = &be_lun->params;
1633 
1634 	be_lun->dev_type = CTL_BE_BLOCK_DEV;
1635 	be_lun->backend.dev.cdev = be_lun->vn->v_rdev;
1636 	be_lun->backend.dev.csw = dev_refthread(be_lun->backend.dev.cdev,
1637 					     &be_lun->backend.dev.dev_ref);
1638 	if (be_lun->backend.dev.csw == NULL)
1639 		panic("Unable to retrieve device switch");
1640 	if (strcmp(be_lun->backend.dev.csw->d_name, "zvol") == 0)
1641 		be_lun->dispatch = ctl_be_block_dispatch_zvol;
1642 	else
1643 		be_lun->dispatch = ctl_be_block_dispatch_dev;
1644 	be_lun->lun_flush = ctl_be_block_flush_dev;
1645 	be_lun->unmap = ctl_be_block_unmap_dev;
1646 
1647 	error = VOP_GETATTR(be_lun->vn, &vattr, NOCRED);
1648 	if (error) {
1649 		snprintf(req->error_str, sizeof(req->error_str),
1650 			 "error getting vnode attributes for device %s",
1651 			 be_lun->dev_path);
1652 		return (error);
1653 	}
1654 
1655 	dev = be_lun->vn->v_rdev;
1656 	devsw = dev->si_devsw;
1657 	if (!devsw->d_ioctl) {
1658 		snprintf(req->error_str, sizeof(req->error_str),
1659 			 "no d_ioctl for device %s!",
1660 			 be_lun->dev_path);
1661 		return (ENODEV);
1662 	}
1663 
1664 	error = devsw->d_ioctl(dev, DIOCGSECTORSIZE,
1665 			       (caddr_t)&be_lun->blocksize, FREAD,
1666 			       curthread);
1667 	if (error) {
1668 		snprintf(req->error_str, sizeof(req->error_str),
1669 			 "error %d returned for DIOCGSECTORSIZE ioctl "
1670 			 "on %s!", error, be_lun->dev_path);
1671 		return (error);
1672 	}
1673 
1674 	/*
1675 	 * If the user has asked for a blocksize that is greater than the
1676 	 * backing device's blocksize, we can do it only if the blocksize
1677 	 * the user is asking for is an even multiple of the underlying
1678 	 * device's blocksize.
1679 	 */
1680 	if ((params->blocksize_bytes != 0)
1681 	 && (params->blocksize_bytes > be_lun->blocksize)) {
1682 		uint32_t bs_multiple, tmp_blocksize;
1683 
1684 		bs_multiple = params->blocksize_bytes / be_lun->blocksize;
1685 
1686 		tmp_blocksize = bs_multiple * be_lun->blocksize;
1687 
1688 		if (tmp_blocksize == params->blocksize_bytes) {
1689 			be_lun->blocksize = params->blocksize_bytes;
1690 		} else {
1691 			snprintf(req->error_str, sizeof(req->error_str),
1692 				 "requested blocksize %u is not an even "
1693 				 "multiple of backing device blocksize %u",
1694 				 params->blocksize_bytes,
1695 				 be_lun->blocksize);
1696 			return (EINVAL);
1697 
1698 		}
1699 	} else if ((params->blocksize_bytes != 0)
1700 		&& (params->blocksize_bytes != be_lun->blocksize)) {
1701 		snprintf(req->error_str, sizeof(req->error_str),
1702 			 "requested blocksize %u < backing device "
1703 			 "blocksize %u", params->blocksize_bytes,
1704 			 be_lun->blocksize);
1705 		return (EINVAL);
1706 	}
1707 
1708 	error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
1709 			       (caddr_t)&be_lun->size_bytes, FREAD,
1710 			       curthread);
1711 	if (error) {
1712 		snprintf(req->error_str, sizeof(req->error_str),
1713 			 "error %d returned for DIOCGMEDIASIZE "
1714 			 " ioctl on %s!", error,
1715 			 be_lun->dev_path);
1716 		return (error);
1717 	}
1718 
1719 	if (params->lun_size_bytes != 0) {
1720 		if (params->lun_size_bytes > be_lun->size_bytes) {
1721 			snprintf(req->error_str, sizeof(req->error_str),
1722 				 "requested LUN size %ju > backing device "
1723 				 "size %ju",
1724 				 (uintmax_t)params->lun_size_bytes,
1725 				 (uintmax_t)be_lun->size_bytes);
1726 			return (EINVAL);
1727 		}
1728 
1729 		be_lun->size_bytes = params->lun_size_bytes;
1730 	}
1731 
1732 	error = devsw->d_ioctl(dev, DIOCGSTRIPESIZE,
1733 			       (caddr_t)&ps, FREAD, curthread);
1734 	if (error)
1735 		ps = po = 0;
1736 	else {
1737 		error = devsw->d_ioctl(dev, DIOCGSTRIPEOFFSET,
1738 				       (caddr_t)&po, FREAD, curthread);
1739 		if (error)
1740 			po = 0;
1741 	}
1742 	pss = ps / be_lun->blocksize;
1743 	pos = po / be_lun->blocksize;
1744 	if ((pss > 0) && (pss * be_lun->blocksize == ps) && (pss >= pos) &&
1745 	    ((pss & (pss - 1)) == 0) && (pos * be_lun->blocksize == po)) {
1746 		be_lun->pblockexp = fls(pss) - 1;
1747 		be_lun->pblockoff = (pss - pos) % pss;
1748 	}
1749 
1750 	return (0);
1751 }
1752 
1753 static int
1754 ctl_be_block_close(struct ctl_be_block_lun *be_lun)
1755 {
1756 	DROP_GIANT();
1757 	if (be_lun->vn) {
1758 		int flags = FREAD | FWRITE;
1759 
1760 		switch (be_lun->dev_type) {
1761 		case CTL_BE_BLOCK_DEV:
1762 			if (be_lun->backend.dev.csw) {
1763 				dev_relthread(be_lun->backend.dev.cdev,
1764 					      be_lun->backend.dev.dev_ref);
1765 				be_lun->backend.dev.csw  = NULL;
1766 				be_lun->backend.dev.cdev = NULL;
1767 			}
1768 			break;
1769 		case CTL_BE_BLOCK_FILE:
1770 			break;
1771 		case CTL_BE_BLOCK_NONE:
1772 			break;
1773 		default:
1774 			panic("Unexpected backend type.");
1775 			break;
1776 		}
1777 
1778 		(void)vn_close(be_lun->vn, flags, NOCRED, curthread);
1779 		be_lun->vn = NULL;
1780 
1781 		switch (be_lun->dev_type) {
1782 		case CTL_BE_BLOCK_DEV:
1783 			break;
1784 		case CTL_BE_BLOCK_FILE:
1785 			if (be_lun->backend.file.cred != NULL) {
1786 				crfree(be_lun->backend.file.cred);
1787 				be_lun->backend.file.cred = NULL;
1788 			}
1789 			break;
1790 		case CTL_BE_BLOCK_NONE:
1791 			break;
1792 		default:
1793 			panic("Unexpected backend type.");
1794 			break;
1795 		}
1796 		be_lun->dev_type = CTL_BE_BLOCK_NONE;
1797 	}
1798 	PICKUP_GIANT();
1799 
1800 	return (0);
1801 }
1802 
1803 static int
1804 ctl_be_block_open(struct ctl_be_block_softc *softc,
1805 		       struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1806 {
1807 	struct nameidata nd;
1808 	int		 flags;
1809 	int		 error;
1810 
1811 	/*
1812 	 * XXX KDM allow a read-only option?
1813 	 */
1814 	flags = FREAD | FWRITE;
1815 	error = 0;
1816 
1817 	if (rootvnode == NULL) {
1818 		snprintf(req->error_str, sizeof(req->error_str),
1819 			 "Root filesystem is not mounted");
1820 		return (1);
1821 	}
1822 
1823 	if (!curthread->td_proc->p_fd->fd_cdir) {
1824 		curthread->td_proc->p_fd->fd_cdir = rootvnode;
1825 		VREF(rootvnode);
1826 	}
1827 	if (!curthread->td_proc->p_fd->fd_rdir) {
1828 		curthread->td_proc->p_fd->fd_rdir = rootvnode;
1829 		VREF(rootvnode);
1830 	}
1831 	if (!curthread->td_proc->p_fd->fd_jdir) {
1832 		curthread->td_proc->p_fd->fd_jdir = rootvnode;
1833 		VREF(rootvnode);
1834 	}
1835 
1836  again:
1837 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, be_lun->dev_path, curthread);
1838 	error = vn_open(&nd, &flags, 0, NULL);
1839 	if (error) {
1840 		/*
1841 		 * This is the only reasonable guess we can make as far as
1842 		 * path if the user doesn't give us a fully qualified path.
1843 		 * If they want to specify a file, they need to specify the
1844 		 * full path.
1845 		 */
1846 		if (be_lun->dev_path[0] != '/') {
1847 			char *dev_path = "/dev/";
1848 			char *dev_name;
1849 
1850 			/* Try adding device path at beginning of name */
1851 			dev_name = malloc(strlen(be_lun->dev_path)
1852 					+ strlen(dev_path) + 1,
1853 					  M_CTLBLK, M_WAITOK);
1854 			if (dev_name) {
1855 				sprintf(dev_name, "%s%s", dev_path,
1856 					be_lun->dev_path);
1857 				free(be_lun->dev_path, M_CTLBLK);
1858 				be_lun->dev_path = dev_name;
1859 				goto again;
1860 			}
1861 		}
1862 		snprintf(req->error_str, sizeof(req->error_str),
1863 		    "error opening %s: %d", be_lun->dev_path, error);
1864 		return (error);
1865 	}
1866 
1867 	NDFREE(&nd, NDF_ONLY_PNBUF);
1868 
1869 	be_lun->vn = nd.ni_vp;
1870 
1871 	/* We only support disks and files. */
1872 	if (vn_isdisk(be_lun->vn, &error)) {
1873 		error = ctl_be_block_open_dev(be_lun, req);
1874 	} else if (be_lun->vn->v_type == VREG) {
1875 		error = ctl_be_block_open_file(be_lun, req);
1876 	} else {
1877 		error = EINVAL;
1878 		snprintf(req->error_str, sizeof(req->error_str),
1879 			 "%s is not a disk or plain file", be_lun->dev_path);
1880 	}
1881 	VOP_UNLOCK(be_lun->vn, 0);
1882 
1883 	if (error != 0) {
1884 		ctl_be_block_close(be_lun);
1885 		return (error);
1886 	}
1887 
1888 	be_lun->blocksize_shift = fls(be_lun->blocksize) - 1;
1889 	be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
1890 
1891 	return (0);
1892 }
1893 
1894 static int
1895 ctl_be_block_create(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
1896 {
1897 	struct ctl_be_block_lun *be_lun;
1898 	struct ctl_lun_create_params *params;
1899 	char num_thread_str[16];
1900 	char tmpstr[32];
1901 	char *value;
1902 	int retval, num_threads, unmap;
1903 	int tmp_num_threads;
1904 
1905 	params = &req->reqdata.create;
1906 	retval = 0;
1907 	req->status = CTL_LUN_OK;
1908 
1909 	num_threads = cbb_num_threads;
1910 
1911 	be_lun = malloc(sizeof(*be_lun), M_CTLBLK, M_ZERO | M_WAITOK);
1912 
1913 	be_lun->params = req->reqdata.create;
1914 	be_lun->softc = softc;
1915 	STAILQ_INIT(&be_lun->input_queue);
1916 	STAILQ_INIT(&be_lun->config_write_queue);
1917 	STAILQ_INIT(&be_lun->datamove_queue);
1918 	sprintf(be_lun->lunname, "cblk%d", softc->num_luns);
1919 	mtx_init(&be_lun->io_lock, "cblk io lock", NULL, MTX_DEF);
1920 	mtx_init(&be_lun->queue_lock, "cblk queue lock", NULL, MTX_DEF);
1921 	ctl_init_opts(&be_lun->ctl_be_lun.options,
1922 	    req->num_be_args, req->kern_be_args);
1923 
1924 	be_lun->lun_zone = uma_zcreate(be_lun->lunname, CTLBLK_MAX_SEG,
1925 	    NULL, NULL, NULL, NULL, /*align*/ 0, /*flags*/0);
1926 
1927 	if (be_lun->lun_zone == NULL) {
1928 		snprintf(req->error_str, sizeof(req->error_str),
1929 			 "error allocating UMA zone");
1930 		goto bailout_error;
1931 	}
1932 
1933 	if (params->flags & CTL_LUN_FLAG_DEV_TYPE)
1934 		be_lun->ctl_be_lun.lun_type = params->device_type;
1935 	else
1936 		be_lun->ctl_be_lun.lun_type = T_DIRECT;
1937 
1938 	if (be_lun->ctl_be_lun.lun_type == T_DIRECT) {
1939 		value = ctl_get_opt(&be_lun->ctl_be_lun.options, "file");
1940 		if (value == NULL) {
1941 			snprintf(req->error_str, sizeof(req->error_str),
1942 				 "no file argument specified");
1943 			goto bailout_error;
1944 		}
1945 		be_lun->dev_path = strdup(value, M_CTLBLK);
1946 		be_lun->blocksize = 512;
1947 		be_lun->blocksize_shift = fls(be_lun->blocksize) - 1;
1948 
1949 		retval = ctl_be_block_open(softc, be_lun, req);
1950 		if (retval != 0) {
1951 			retval = 0;
1952 			req->status = CTL_LUN_WARNING;
1953 		}
1954 	} else {
1955 		/*
1956 		 * For processor devices, we don't have any size.
1957 		 */
1958 		be_lun->blocksize = 0;
1959 		be_lun->pblockexp = 0;
1960 		be_lun->pblockoff = 0;
1961 		be_lun->size_blocks = 0;
1962 		be_lun->size_bytes = 0;
1963 		be_lun->ctl_be_lun.maxlba = 0;
1964 
1965 		/*
1966 		 * Default to just 1 thread for processor devices.
1967 		 */
1968 		num_threads = 1;
1969 	}
1970 
1971 	/*
1972 	 * XXX This searching loop might be refactored to be combined with
1973 	 * the loop above,
1974 	 */
1975 	value = ctl_get_opt(&be_lun->ctl_be_lun.options, "num_threads");
1976 	if (value != NULL) {
1977 		tmp_num_threads = strtol(value, NULL, 0);
1978 
1979 		/*
1980 		 * We don't let the user specify less than one
1981 		 * thread, but hope he's clueful enough not to
1982 		 * specify 1000 threads.
1983 		 */
1984 		if (tmp_num_threads < 1) {
1985 			snprintf(req->error_str, sizeof(req->error_str),
1986 				 "invalid number of threads %s",
1987 				 num_thread_str);
1988 			goto bailout_error;
1989 		}
1990 		num_threads = tmp_num_threads;
1991 	}
1992 	unmap = 0;
1993 	value = ctl_get_opt(&be_lun->ctl_be_lun.options, "unmap");
1994 	if (value != NULL && strcmp(value, "on") == 0)
1995 		unmap = 1;
1996 
1997 	be_lun->flags = CTL_BE_BLOCK_LUN_UNCONFIGURED;
1998 	be_lun->ctl_be_lun.flags = CTL_LUN_FLAG_PRIMARY;
1999 	if (be_lun->vn == NULL)
2000 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_OFFLINE;
2001 	if (unmap)
2002 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_UNMAP;
2003 	be_lun->ctl_be_lun.be_lun = be_lun;
2004 	be_lun->ctl_be_lun.maxlba = (be_lun->size_blocks == 0) ?
2005 	    0 : (be_lun->size_blocks - 1);
2006 	be_lun->ctl_be_lun.blocksize = be_lun->blocksize;
2007 	be_lun->ctl_be_lun.pblockexp = be_lun->pblockexp;
2008 	be_lun->ctl_be_lun.pblockoff = be_lun->pblockoff;
2009 	if (be_lun->dispatch == ctl_be_block_dispatch_zvol &&
2010 	    be_lun->blocksize != 0)
2011 		be_lun->ctl_be_lun.atomicblock = CTLBLK_MAX_IO_SIZE /
2012 		    be_lun->blocksize;
2013 	/* Tell the user the blocksize we ended up using */
2014 	params->lun_size_bytes = be_lun->size_bytes;
2015 	params->blocksize_bytes = be_lun->blocksize;
2016 	if (params->flags & CTL_LUN_FLAG_ID_REQ) {
2017 		be_lun->ctl_be_lun.req_lun_id = params->req_lun_id;
2018 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_ID_REQ;
2019 	} else
2020 		be_lun->ctl_be_lun.req_lun_id = 0;
2021 
2022 	be_lun->ctl_be_lun.lun_shutdown = ctl_be_block_lun_shutdown;
2023 	be_lun->ctl_be_lun.lun_config_status =
2024 		ctl_be_block_lun_config_status;
2025 	be_lun->ctl_be_lun.be = &ctl_be_block_driver;
2026 
2027 	if ((params->flags & CTL_LUN_FLAG_SERIAL_NUM) == 0) {
2028 		snprintf(tmpstr, sizeof(tmpstr), "MYSERIAL%4d",
2029 			 softc->num_luns);
2030 		strncpy((char *)be_lun->ctl_be_lun.serial_num, tmpstr,
2031 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
2032 			sizeof(tmpstr)));
2033 
2034 		/* Tell the user what we used for a serial number */
2035 		strncpy((char *)params->serial_num, tmpstr,
2036 			ctl_min(sizeof(params->serial_num), sizeof(tmpstr)));
2037 	} else {
2038 		strncpy((char *)be_lun->ctl_be_lun.serial_num,
2039 			params->serial_num,
2040 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
2041 			sizeof(params->serial_num)));
2042 	}
2043 	if ((params->flags & CTL_LUN_FLAG_DEVID) == 0) {
2044 		snprintf(tmpstr, sizeof(tmpstr), "MYDEVID%4d", softc->num_luns);
2045 		strncpy((char *)be_lun->ctl_be_lun.device_id, tmpstr,
2046 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
2047 			sizeof(tmpstr)));
2048 
2049 		/* Tell the user what we used for a device ID */
2050 		strncpy((char *)params->device_id, tmpstr,
2051 			ctl_min(sizeof(params->device_id), sizeof(tmpstr)));
2052 	} else {
2053 		strncpy((char *)be_lun->ctl_be_lun.device_id,
2054 			params->device_id,
2055 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
2056 				sizeof(params->device_id)));
2057 	}
2058 
2059 	TASK_INIT(&be_lun->io_task, /*priority*/0, ctl_be_block_worker, be_lun);
2060 
2061 	be_lun->io_taskqueue = taskqueue_create(be_lun->lunname, M_WAITOK,
2062 	    taskqueue_thread_enqueue, /*context*/&be_lun->io_taskqueue);
2063 
2064 	if (be_lun->io_taskqueue == NULL) {
2065 		snprintf(req->error_str, sizeof(req->error_str),
2066 			 "unable to create taskqueue");
2067 		goto bailout_error;
2068 	}
2069 
2070 	/*
2071 	 * Note that we start the same number of threads by default for
2072 	 * both the file case and the block device case.  For the file
2073 	 * case, we need multiple threads to allow concurrency, because the
2074 	 * vnode interface is designed to be a blocking interface.  For the
2075 	 * block device case, ZFS zvols at least will block the caller's
2076 	 * context in many instances, and so we need multiple threads to
2077 	 * overcome that problem.  Other block devices don't need as many
2078 	 * threads, but they shouldn't cause too many problems.
2079 	 *
2080 	 * If the user wants to just have a single thread for a block
2081 	 * device, he can specify that when the LUN is created, or change
2082 	 * the tunable/sysctl to alter the default number of threads.
2083 	 */
2084 	retval = taskqueue_start_threads(&be_lun->io_taskqueue,
2085 					 /*num threads*/num_threads,
2086 					 /*priority*/PWAIT,
2087 					 /*thread name*/
2088 					 "%s taskq", be_lun->lunname);
2089 
2090 	if (retval != 0)
2091 		goto bailout_error;
2092 
2093 	be_lun->num_threads = num_threads;
2094 
2095 	mtx_lock(&softc->lock);
2096 	softc->num_luns++;
2097 	STAILQ_INSERT_TAIL(&softc->lun_list, be_lun, links);
2098 
2099 	mtx_unlock(&softc->lock);
2100 
2101 	retval = ctl_add_lun(&be_lun->ctl_be_lun);
2102 	if (retval != 0) {
2103 		mtx_lock(&softc->lock);
2104 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2105 			      links);
2106 		softc->num_luns--;
2107 		mtx_unlock(&softc->lock);
2108 		snprintf(req->error_str, sizeof(req->error_str),
2109 			 "ctl_add_lun() returned error %d, see dmesg for "
2110 			 "details", retval);
2111 		retval = 0;
2112 		goto bailout_error;
2113 	}
2114 
2115 	mtx_lock(&softc->lock);
2116 
2117 	/*
2118 	 * Tell the config_status routine that we're waiting so it won't
2119 	 * clean up the LUN in the event of an error.
2120 	 */
2121 	be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
2122 
2123 	while (be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) {
2124 		retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
2125 		if (retval == EINTR)
2126 			break;
2127 	}
2128 	be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
2129 
2130 	if (be_lun->flags & CTL_BE_BLOCK_LUN_CONFIG_ERR) {
2131 		snprintf(req->error_str, sizeof(req->error_str),
2132 			 "LUN configuration error, see dmesg for details");
2133 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2134 			      links);
2135 		softc->num_luns--;
2136 		mtx_unlock(&softc->lock);
2137 		goto bailout_error;
2138 	} else {
2139 		params->req_lun_id = be_lun->ctl_be_lun.lun_id;
2140 	}
2141 
2142 	mtx_unlock(&softc->lock);
2143 
2144 	be_lun->disk_stats = devstat_new_entry("cbb", params->req_lun_id,
2145 					       be_lun->blocksize,
2146 					       DEVSTAT_ALL_SUPPORTED,
2147 					       be_lun->ctl_be_lun.lun_type
2148 					       | DEVSTAT_TYPE_IF_OTHER,
2149 					       DEVSTAT_PRIORITY_OTHER);
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 			 "LUN %u is not managed by the block backend",
2193 			 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 			 "error %d returned from ctl_disable_lun() for "
2202 			 "LUN %d", 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 			 "error %d returned from ctl_invalidate_lun() for "
2211 			 "LUN %d", 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 			 "interrupted waiting for LUN to be freed");
2230 		mtx_unlock(&softc->lock);
2231 		goto bailout_error;
2232 	}
2233 
2234 	STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun, links);
2235 
2236 	softc->num_luns--;
2237 	mtx_unlock(&softc->lock);
2238 
2239 	taskqueue_drain(be_lun->io_taskqueue, &be_lun->io_task);
2240 
2241 	taskqueue_free(be_lun->io_taskqueue);
2242 
2243 	ctl_be_block_close(be_lun);
2244 
2245 	if (be_lun->disk_stats != NULL)
2246 		devstat_remove_entry(be_lun->disk_stats);
2247 
2248 	uma_zdestroy(be_lun->lun_zone);
2249 
2250 	ctl_free_opts(&be_lun->ctl_be_lun.options);
2251 	free(be_lun->dev_path, M_CTLBLK);
2252 	mtx_destroy(&be_lun->queue_lock);
2253 	mtx_destroy(&be_lun->io_lock);
2254 	free(be_lun, M_CTLBLK);
2255 
2256 	req->status = CTL_LUN_OK;
2257 
2258 	return (0);
2259 
2260 bailout_error:
2261 
2262 	req->status = CTL_LUN_ERROR;
2263 
2264 	return (0);
2265 }
2266 
2267 static int
2268 ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
2269 			 struct ctl_lun_req *req)
2270 {
2271 	struct vattr vattr;
2272 	int error;
2273 	struct ctl_lun_create_params *params = &be_lun->params;
2274 
2275 	if (params->lun_size_bytes != 0) {
2276 		be_lun->size_bytes = params->lun_size_bytes;
2277 	} else  {
2278 		vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
2279 		error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
2280 		VOP_UNLOCK(be_lun->vn, 0);
2281 		if (error != 0) {
2282 			snprintf(req->error_str, sizeof(req->error_str),
2283 				 "error calling VOP_GETATTR() for file %s",
2284 				 be_lun->dev_path);
2285 			return (error);
2286 		}
2287 
2288 		be_lun->size_bytes = vattr.va_size;
2289 	}
2290 
2291 	return (0);
2292 }
2293 
2294 static int
2295 ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
2296 			struct ctl_lun_req *req)
2297 {
2298 	struct ctl_be_block_devdata *dev_data;
2299 	int error;
2300 	struct ctl_lun_create_params *params = &be_lun->params;
2301 	uint64_t size_bytes;
2302 
2303 	dev_data = &be_lun->backend.dev;
2304 	if (!dev_data->csw->d_ioctl) {
2305 		snprintf(req->error_str, sizeof(req->error_str),
2306 			 "no d_ioctl for device %s!", be_lun->dev_path);
2307 		return (ENODEV);
2308 	}
2309 
2310 	error = dev_data->csw->d_ioctl(dev_data->cdev, DIOCGMEDIASIZE,
2311 			       (caddr_t)&size_bytes, FREAD,
2312 			       curthread);
2313 	if (error) {
2314 		snprintf(req->error_str, sizeof(req->error_str),
2315 			 "error %d returned for DIOCGMEDIASIZE ioctl "
2316 			 "on %s!", error, be_lun->dev_path);
2317 		return (error);
2318 	}
2319 
2320 	if (params->lun_size_bytes != 0) {
2321 		if (params->lun_size_bytes > size_bytes) {
2322 			snprintf(req->error_str, sizeof(req->error_str),
2323 				 "requested LUN size %ju > backing device "
2324 				 "size %ju",
2325 				 (uintmax_t)params->lun_size_bytes,
2326 				 (uintmax_t)size_bytes);
2327 			return (EINVAL);
2328 		}
2329 
2330 		be_lun->size_bytes = params->lun_size_bytes;
2331 	} else {
2332 		be_lun->size_bytes = size_bytes;
2333 	}
2334 
2335 	return (0);
2336 }
2337 
2338 static int
2339 ctl_be_block_modify(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2340 {
2341 	struct ctl_lun_modify_params *params;
2342 	struct ctl_be_block_lun *be_lun;
2343 	uint64_t oldsize;
2344 	int error;
2345 
2346 	params = &req->reqdata.modify;
2347 
2348 	mtx_lock(&softc->lock);
2349 	be_lun = NULL;
2350 	STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2351 		if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2352 			break;
2353 	}
2354 	mtx_unlock(&softc->lock);
2355 
2356 	if (be_lun == NULL) {
2357 		snprintf(req->error_str, sizeof(req->error_str),
2358 			 "LUN %u is not managed by the block backend",
2359 			 params->lun_id);
2360 		goto bailout_error;
2361 	}
2362 
2363 	be_lun->params.lun_size_bytes = params->lun_size_bytes;
2364 
2365 	oldsize = be_lun->size_blocks;
2366 	if (be_lun->vn == NULL)
2367 		error = ctl_be_block_open(softc, be_lun, req);
2368 	else if (be_lun->vn->v_type == VREG)
2369 		error = ctl_be_block_modify_file(be_lun, req);
2370 	else
2371 		error = ctl_be_block_modify_dev(be_lun, req);
2372 
2373 	if (error == 0 && be_lun->size_blocks != oldsize) {
2374 		be_lun->size_blocks = be_lun->size_bytes >>
2375 		    be_lun->blocksize_shift;
2376 
2377 		/*
2378 		 * The maximum LBA is the size - 1.
2379 		 *
2380 		 * XXX: Note that this field is being updated without locking,
2381 		 * 	which might cause problems on 32-bit architectures.
2382 		 */
2383 		be_lun->ctl_be_lun.maxlba = (be_lun->size_blocks == 0) ?
2384 		    0 : (be_lun->size_blocks - 1);
2385 		be_lun->ctl_be_lun.blocksize = be_lun->blocksize;
2386 		be_lun->ctl_be_lun.pblockexp = be_lun->pblockexp;
2387 		be_lun->ctl_be_lun.pblockoff = be_lun->pblockoff;
2388 		if (be_lun->dispatch == ctl_be_block_dispatch_zvol &&
2389 		    be_lun->blocksize != 0)
2390 			be_lun->ctl_be_lun.atomicblock = CTLBLK_MAX_IO_SIZE /
2391 			    be_lun->blocksize;
2392 		ctl_lun_capacity_changed(&be_lun->ctl_be_lun);
2393 		if (oldsize == 0 && be_lun->size_blocks != 0)
2394 			ctl_lun_online(&be_lun->ctl_be_lun);
2395 	}
2396 
2397 	/* Tell the user the exact size we ended up using */
2398 	params->lun_size_bytes = be_lun->size_bytes;
2399 
2400 	req->status = error ? CTL_LUN_WARNING : CTL_LUN_OK;
2401 
2402 	return (0);
2403 
2404 bailout_error:
2405 	req->status = CTL_LUN_ERROR;
2406 
2407 	return (0);
2408 }
2409 
2410 static void
2411 ctl_be_block_lun_shutdown(void *be_lun)
2412 {
2413 	struct ctl_be_block_lun *lun;
2414 	struct ctl_be_block_softc *softc;
2415 
2416 	lun = (struct ctl_be_block_lun *)be_lun;
2417 
2418 	softc = lun->softc;
2419 
2420 	mtx_lock(&softc->lock);
2421 	lun->flags |= CTL_BE_BLOCK_LUN_UNCONFIGURED;
2422 	if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2423 		wakeup(lun);
2424 	mtx_unlock(&softc->lock);
2425 
2426 }
2427 
2428 static void
2429 ctl_be_block_lun_config_status(void *be_lun, ctl_lun_config_status status)
2430 {
2431 	struct ctl_be_block_lun *lun;
2432 	struct ctl_be_block_softc *softc;
2433 
2434 	lun = (struct ctl_be_block_lun *)be_lun;
2435 	softc = lun->softc;
2436 
2437 	if (status == CTL_LUN_CONFIG_OK) {
2438 		mtx_lock(&softc->lock);
2439 		lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2440 		if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2441 			wakeup(lun);
2442 		mtx_unlock(&softc->lock);
2443 
2444 		/*
2445 		 * We successfully added the LUN, attempt to enable it.
2446 		 */
2447 		if (ctl_enable_lun(&lun->ctl_be_lun) != 0) {
2448 			printf("%s: ctl_enable_lun() failed!\n", __func__);
2449 			if (ctl_invalidate_lun(&lun->ctl_be_lun) != 0) {
2450 				printf("%s: ctl_invalidate_lun() failed!\n",
2451 				       __func__);
2452 			}
2453 		}
2454 
2455 		return;
2456 	}
2457 
2458 
2459 	mtx_lock(&softc->lock);
2460 	lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2461 	lun->flags |= CTL_BE_BLOCK_LUN_CONFIG_ERR;
2462 	wakeup(lun);
2463 	mtx_unlock(&softc->lock);
2464 }
2465 
2466 
2467 static int
2468 ctl_be_block_config_write(union ctl_io *io)
2469 {
2470 	struct ctl_be_block_lun *be_lun;
2471 	struct ctl_be_lun *ctl_be_lun;
2472 	int retval;
2473 
2474 	retval = 0;
2475 
2476 	DPRINTF("entered\n");
2477 
2478 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
2479 		CTL_PRIV_BACKEND_LUN].ptr;
2480 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
2481 
2482 	switch (io->scsiio.cdb[0]) {
2483 	case SYNCHRONIZE_CACHE:
2484 	case SYNCHRONIZE_CACHE_16:
2485 	case WRITE_SAME_10:
2486 	case WRITE_SAME_16:
2487 	case UNMAP:
2488 		/*
2489 		 * The upper level CTL code will filter out any CDBs with
2490 		 * the immediate bit set and return the proper error.
2491 		 *
2492 		 * We don't really need to worry about what LBA range the
2493 		 * user asked to be synced out.  When they issue a sync
2494 		 * cache command, we'll sync out the whole thing.
2495 		 */
2496 		mtx_lock(&be_lun->queue_lock);
2497 		STAILQ_INSERT_TAIL(&be_lun->config_write_queue, &io->io_hdr,
2498 				   links);
2499 		mtx_unlock(&be_lun->queue_lock);
2500 		taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
2501 		break;
2502 	case START_STOP_UNIT: {
2503 		struct scsi_start_stop_unit *cdb;
2504 
2505 		cdb = (struct scsi_start_stop_unit *)io->scsiio.cdb;
2506 
2507 		if (cdb->how & SSS_START)
2508 			retval = ctl_start_lun(ctl_be_lun);
2509 		else {
2510 			retval = ctl_stop_lun(ctl_be_lun);
2511 			/*
2512 			 * XXX KDM Copan-specific offline behavior.
2513 			 * Figure out a reasonable way to port this?
2514 			 */
2515 #ifdef NEEDTOPORT
2516 			if ((retval == 0)
2517 			 && (cdb->byte2 & SSS_ONOFFLINE))
2518 				retval = ctl_lun_offline(ctl_be_lun);
2519 #endif
2520 		}
2521 
2522 		/*
2523 		 * In general, the above routines should not fail.  They
2524 		 * just set state for the LUN.  So we've got something
2525 		 * pretty wrong here if we can't start or stop the LUN.
2526 		 */
2527 		if (retval != 0) {
2528 			ctl_set_internal_failure(&io->scsiio,
2529 						 /*sks_valid*/ 1,
2530 						 /*retry_count*/ 0xf051);
2531 			retval = CTL_RETVAL_COMPLETE;
2532 		} else {
2533 			ctl_set_success(&io->scsiio);
2534 		}
2535 		ctl_config_write_done(io);
2536 		break;
2537 	}
2538 	default:
2539 		ctl_set_invalid_opcode(&io->scsiio);
2540 		ctl_config_write_done(io);
2541 		retval = CTL_RETVAL_COMPLETE;
2542 		break;
2543 	}
2544 
2545 	return (retval);
2546 
2547 }
2548 
2549 static int
2550 ctl_be_block_config_read(union ctl_io *io)
2551 {
2552 	return (0);
2553 }
2554 
2555 static int
2556 ctl_be_block_lun_info(void *be_lun, struct sbuf *sb)
2557 {
2558 	struct ctl_be_block_lun *lun;
2559 	int retval;
2560 
2561 	lun = (struct ctl_be_block_lun *)be_lun;
2562 	retval = 0;
2563 
2564 	retval = sbuf_printf(sb, "\t<num_threads>");
2565 
2566 	if (retval != 0)
2567 		goto bailout;
2568 
2569 	retval = sbuf_printf(sb, "%d", lun->num_threads);
2570 
2571 	if (retval != 0)
2572 		goto bailout;
2573 
2574 	retval = sbuf_printf(sb, "</num_threads>\n");
2575 
2576 bailout:
2577 
2578 	return (retval);
2579 }
2580 
2581 int
2582 ctl_be_block_init(void)
2583 {
2584 	struct ctl_be_block_softc *softc;
2585 	int retval;
2586 
2587 	softc = &backend_block_softc;
2588 	retval = 0;
2589 
2590 	mtx_init(&softc->lock, "ctlblock", NULL, MTX_DEF);
2591 	beio_zone = uma_zcreate("beio", sizeof(struct ctl_be_block_io),
2592 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
2593 	STAILQ_INIT(&softc->disk_list);
2594 	STAILQ_INIT(&softc->lun_list);
2595 
2596 	return (retval);
2597 }
2598