xref: /freebsd/sys/cam/ctl/ctl_backend_block.c (revision 95d45410b5100e07f6f98450bcd841a8945d4726)
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				bio_flags;
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 = beio->bio_flags;
603 
604 	bzero(&xuio, sizeof(xuio));
605 	if (beio->bio_cmd == BIO_READ) {
606 		SDT_PROBE(cbb, kernel, read, file_start, 0, 0, 0, 0, 0);
607 		xuio.uio_rw = UIO_READ;
608 	} else {
609 		SDT_PROBE(cbb, kernel, write, file_start, 0, 0, 0, 0, 0);
610 		xuio.uio_rw = UIO_WRITE;
611 	}
612 	xuio.uio_offset = beio->io_offset;
613 	xuio.uio_resid = beio->io_len;
614 	xuio.uio_segflg = UIO_SYSSPACE;
615 	xuio.uio_iov = beio->xiovecs;
616 	xuio.uio_iovcnt = beio->num_segs;
617 	xuio.uio_td = curthread;
618 
619 	for (i = 0, xiovec = xuio.uio_iov; i < xuio.uio_iovcnt; i++, xiovec++) {
620 		xiovec->iov_base = beio->sg_segs[i].addr;
621 		xiovec->iov_len = beio->sg_segs[i].len;
622 	}
623 
624 	binuptime(&beio->ds_t0);
625 	mtx_lock(&be_lun->io_lock);
626 	devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
627 	mtx_unlock(&be_lun->io_lock);
628 
629 	if (beio->bio_cmd == BIO_READ) {
630 		vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
631 
632 		/*
633 		 * UFS pays attention to IO_DIRECT for reads.  If the
634 		 * DIRECTIO option is configured into the kernel, it calls
635 		 * ffs_rawread().  But that only works for single-segment
636 		 * uios with user space addresses.  In our case, with a
637 		 * kernel uio, it still reads into the buffer cache, but it
638 		 * will just try to release the buffer from the cache later
639 		 * on in ffs_read().
640 		 *
641 		 * ZFS does not pay attention to IO_DIRECT for reads.
642 		 *
643 		 * UFS does not pay attention to IO_SYNC for reads.
644 		 *
645 		 * ZFS pays attention to IO_SYNC (which translates into the
646 		 * Solaris define FRSYNC for zfs_read()) for reads.  It
647 		 * attempts to sync the file before reading.
648 		 *
649 		 * So, to attempt to provide some barrier semantics in the
650 		 * BIO_ORDERED case, set both IO_DIRECT and IO_SYNC.
651 		 */
652 		error = VOP_READ(be_lun->vn, &xuio, (flags & BIO_ORDERED) ?
653 				 (IO_DIRECT|IO_SYNC) : 0, file_data->cred);
654 
655 		VOP_UNLOCK(be_lun->vn, 0);
656 		SDT_PROBE(cbb, kernel, read, file_done, 0, 0, 0, 0, 0);
657 	} else {
658 		struct mount *mountpoint;
659 		int lock_flags;
660 
661 		(void)vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
662 
663 		if (MNT_SHARED_WRITES(mountpoint)
664 		 || ((mountpoint == NULL)
665 		  && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
666 			lock_flags = LK_SHARED;
667 		else
668 			lock_flags = LK_EXCLUSIVE;
669 
670 		vn_lock(be_lun->vn, lock_flags | LK_RETRY);
671 
672 		/*
673 		 * UFS pays attention to IO_DIRECT for writes.  The write
674 		 * is done asynchronously.  (Normally the write would just
675 		 * get put into cache.
676 		 *
677 		 * UFS pays attention to IO_SYNC for writes.  It will
678 		 * attempt to write the buffer out synchronously if that
679 		 * flag is set.
680 		 *
681 		 * ZFS does not pay attention to IO_DIRECT for writes.
682 		 *
683 		 * ZFS pays attention to IO_SYNC (a.k.a. FSYNC or FRSYNC)
684 		 * for writes.  It will flush the transaction from the
685 		 * cache before returning.
686 		 *
687 		 * So if we've got the BIO_ORDERED flag set, we want
688 		 * IO_SYNC in either the UFS or ZFS case.
689 		 */
690 		error = VOP_WRITE(be_lun->vn, &xuio, (flags & BIO_ORDERED) ?
691 				  IO_SYNC : 0, file_data->cred);
692 		VOP_UNLOCK(be_lun->vn, 0);
693 
694 		vn_finished_write(mountpoint);
695 		SDT_PROBE(cbb, kernel, write, file_done, 0, 0, 0, 0, 0);
696         }
697 
698 	mtx_lock(&be_lun->io_lock);
699 	devstat_end_transaction(beio->lun->disk_stats, beio->io_len,
700 	    beio->ds_tag_type, beio->ds_trans_type,
701 	    /*now*/ NULL, /*then*/&beio->ds_t0);
702 	mtx_unlock(&be_lun->io_lock);
703 
704 	/*
705 	 * If we got an error, set the sense data to "MEDIUM ERROR" and
706 	 * return the I/O to the user.
707 	 */
708 	if (error != 0) {
709 		char path_str[32];
710 
711 		ctl_scsi_path_string(io, path_str, sizeof(path_str));
712 		/*
713 		 * XXX KDM ZFS returns ENOSPC when the underlying
714 		 * filesystem fills up.  What kind of SCSI error should we
715 		 * return for that?
716 		 */
717 		printf("%s%s command returned errno %d\n", path_str,
718 		       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE", error);
719 		ctl_set_medium_error(&io->scsiio);
720 		ctl_complete_beio(beio);
721 		return;
722 	}
723 
724 	/*
725 	 * If this is a write, we're all done.
726 	 * If this is a read, we can now send the data to the user.
727 	 */
728 	if (ARGS(io)->flags & (CTL_LLF_WRITE | CTL_LLF_VERIFY)) {
729 		ctl_set_success(&io->scsiio);
730 		ctl_complete_beio(beio);
731 	} else {
732 #ifdef CTL_TIME_IO
733         	getbintime(&io->io_hdr.dma_start_bt);
734 #endif
735 		ctl_datamove(io);
736 	}
737 }
738 
739 static void
740 ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
741 		       struct ctl_be_block_io *beio)
742 {
743 	struct bio *bio;
744 	union ctl_io *io;
745 	struct ctl_be_block_devdata *dev_data;
746 
747 	dev_data = &be_lun->backend.dev;
748 	io = beio->io;
749 
750 	DPRINTF("entered\n");
751 
752 	/* This can't fail, it's a blocking allocation. */
753 	bio = g_alloc_bio();
754 
755 	bio->bio_cmd	    = BIO_FLUSH;
756 	bio->bio_flags	   |= BIO_ORDERED;
757 	bio->bio_dev	    = dev_data->cdev;
758 	bio->bio_offset	    = 0;
759 	bio->bio_data	    = 0;
760 	bio->bio_done	    = ctl_be_block_biodone;
761 	bio->bio_caller1    = beio;
762 	bio->bio_pblkno	    = 0;
763 
764 	/*
765 	 * We don't need to acquire the LUN lock here, because we are only
766 	 * sending one bio, and so there is no other context to synchronize
767 	 * with.
768 	 */
769 	beio->num_bios_sent = 1;
770 	beio->send_complete = 1;
771 
772 	binuptime(&beio->ds_t0);
773 	mtx_lock(&be_lun->io_lock);
774 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
775 	mtx_unlock(&be_lun->io_lock);
776 
777 	(*dev_data->csw->d_strategy)(bio);
778 }
779 
780 static void
781 ctl_be_block_unmap_dev_range(struct ctl_be_block_lun *be_lun,
782 		       struct ctl_be_block_io *beio,
783 		       uint64_t off, uint64_t len, int last)
784 {
785 	struct bio *bio;
786 	struct ctl_be_block_devdata *dev_data;
787 	uint64_t maxlen;
788 
789 	dev_data = &be_lun->backend.dev;
790 	maxlen = LONG_MAX - (LONG_MAX % be_lun->blocksize);
791 	while (len > 0) {
792 		bio = g_alloc_bio();
793 		bio->bio_cmd	    = BIO_DELETE;
794 		bio->bio_flags	   |= beio->bio_flags;
795 		bio->bio_dev	    = dev_data->cdev;
796 		bio->bio_offset	    = off;
797 		bio->bio_length	    = MIN(len, maxlen);
798 		bio->bio_data	    = 0;
799 		bio->bio_done	    = ctl_be_block_biodone;
800 		bio->bio_caller1    = beio;
801 		bio->bio_pblkno     = off / be_lun->blocksize;
802 
803 		off += bio->bio_length;
804 		len -= bio->bio_length;
805 
806 		mtx_lock(&be_lun->io_lock);
807 		beio->num_bios_sent++;
808 		if (last && len == 0)
809 			beio->send_complete = 1;
810 		mtx_unlock(&be_lun->io_lock);
811 
812 		(*dev_data->csw->d_strategy)(bio);
813 	}
814 }
815 
816 static void
817 ctl_be_block_unmap_dev(struct ctl_be_block_lun *be_lun,
818 		       struct ctl_be_block_io *beio)
819 {
820 	union ctl_io *io;
821 	struct ctl_be_block_devdata *dev_data;
822 	struct ctl_ptr_len_flags *ptrlen;
823 	struct scsi_unmap_desc *buf, *end;
824 	uint64_t len;
825 
826 	dev_data = &be_lun->backend.dev;
827 	io = beio->io;
828 
829 	DPRINTF("entered\n");
830 
831 	binuptime(&beio->ds_t0);
832 	mtx_lock(&be_lun->io_lock);
833 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
834 	mtx_unlock(&be_lun->io_lock);
835 
836 	if (beio->io_offset == -1) {
837 		beio->io_len = 0;
838 		ptrlen = (struct ctl_ptr_len_flags *)&io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN];
839 		buf = (struct scsi_unmap_desc *)ptrlen->ptr;
840 		end = buf + ptrlen->len / sizeof(*buf);
841 		for (; buf < end; buf++) {
842 			len = (uint64_t)scsi_4btoul(buf->length) *
843 			    be_lun->blocksize;
844 			beio->io_len += len;
845 			ctl_be_block_unmap_dev_range(be_lun, beio,
846 			    scsi_8btou64(buf->lba) * be_lun->blocksize, len,
847 			    (end - buf < 2) ? TRUE : FALSE);
848 		}
849 	} else
850 		ctl_be_block_unmap_dev_range(be_lun, beio,
851 		    beio->io_offset, beio->io_len, TRUE);
852 }
853 
854 static void
855 ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
856 			  struct ctl_be_block_io *beio)
857 {
858 	TAILQ_HEAD(, bio) queue = TAILQ_HEAD_INITIALIZER(queue);
859 	int i;
860 	struct bio *bio;
861 	struct ctl_be_block_devdata *dev_data;
862 	off_t cur_offset;
863 	int max_iosize;
864 
865 	DPRINTF("entered\n");
866 
867 	dev_data = &be_lun->backend.dev;
868 
869 	/*
870 	 * We have to limit our I/O size to the maximum supported by the
871 	 * backend device.  Hopefully it is MAXPHYS.  If the driver doesn't
872 	 * set it properly, use DFLTPHYS.
873 	 */
874 	max_iosize = dev_data->cdev->si_iosize_max;
875 	if (max_iosize < PAGE_SIZE)
876 		max_iosize = DFLTPHYS;
877 
878 	cur_offset = beio->io_offset;
879 	for (i = 0; i < beio->num_segs; i++) {
880 		size_t cur_size;
881 		uint8_t *cur_ptr;
882 
883 		cur_size = beio->sg_segs[i].len;
884 		cur_ptr = beio->sg_segs[i].addr;
885 
886 		while (cur_size > 0) {
887 			/* This can't fail, it's a blocking allocation. */
888 			bio = g_alloc_bio();
889 
890 			KASSERT(bio != NULL, ("g_alloc_bio() failed!\n"));
891 
892 			bio->bio_cmd = beio->bio_cmd;
893 			bio->bio_flags |= beio->bio_flags;
894 			bio->bio_dev = dev_data->cdev;
895 			bio->bio_caller1 = beio;
896 			bio->bio_length = min(cur_size, max_iosize);
897 			bio->bio_offset = cur_offset;
898 			bio->bio_data = cur_ptr;
899 			bio->bio_done = ctl_be_block_biodone;
900 			bio->bio_pblkno = cur_offset / be_lun->blocksize;
901 
902 			cur_offset += bio->bio_length;
903 			cur_ptr += bio->bio_length;
904 			cur_size -= bio->bio_length;
905 
906 			TAILQ_INSERT_TAIL(&queue, bio, bio_queue);
907 			beio->num_bios_sent++;
908 		}
909 	}
910 	binuptime(&beio->ds_t0);
911 	mtx_lock(&be_lun->io_lock);
912 	devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
913 	beio->send_complete = 1;
914 	mtx_unlock(&be_lun->io_lock);
915 
916 	/*
917 	 * Fire off all allocated requests!
918 	 */
919 	while ((bio = TAILQ_FIRST(&queue)) != NULL) {
920 		TAILQ_REMOVE(&queue, bio, bio_queue);
921 		(*dev_data->csw->d_strategy)(bio);
922 	}
923 }
924 
925 static void
926 ctl_be_block_cw_done_ws(struct ctl_be_block_io *beio)
927 {
928 	union ctl_io *io;
929 
930 	io = beio->io;
931 	ctl_free_beio(beio);
932 	if ((io->io_hdr.flags & CTL_FLAG_ABORT) ||
933 	    ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE &&
934 	     (io->io_hdr.status & CTL_STATUS_MASK) != CTL_SUCCESS)) {
935 		ctl_config_write_done(io);
936 		return;
937 	}
938 
939 	ctl_be_block_config_write(io);
940 }
941 
942 static void
943 ctl_be_block_cw_dispatch_ws(struct ctl_be_block_lun *be_lun,
944 			    union ctl_io *io)
945 {
946 	struct ctl_be_block_io *beio;
947 	struct ctl_be_block_softc *softc;
948 	struct ctl_lba_len_flags *lbalen;
949 	uint64_t len_left, lba;
950 	int i, seglen;
951 	uint8_t *buf, *end;
952 
953 	DPRINTF("entered\n");
954 
955 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
956 	softc = be_lun->softc;
957 	lbalen = ARGS(beio->io);
958 
959 	if (lbalen->flags & ~(SWS_LBDATA | SWS_UNMAP) ||
960 	    (lbalen->flags & SWS_UNMAP && be_lun->unmap == NULL)) {
961 		ctl_free_beio(beio);
962 		ctl_set_invalid_field(&io->scsiio,
963 				      /*sks_valid*/ 1,
964 				      /*command*/ 1,
965 				      /*field*/ 1,
966 				      /*bit_valid*/ 0,
967 				      /*bit*/ 0);
968 		ctl_config_write_done(io);
969 		return;
970 	}
971 
972 	/*
973 	 * If the I/O came down with an ordered or head of queue tag, set
974 	 * the BIO_ORDERED attribute.  For head of queue tags, that's
975 	 * pretty much the best we can do.
976 	 */
977 	if ((io->scsiio.tag_type == CTL_TAG_ORDERED)
978 	 || (io->scsiio.tag_type == CTL_TAG_HEAD_OF_QUEUE))
979 		beio->bio_flags = BIO_ORDERED;
980 
981 	switch (io->scsiio.tag_type) {
982 	case CTL_TAG_ORDERED:
983 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
984 		break;
985 	case CTL_TAG_HEAD_OF_QUEUE:
986 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
987 		break;
988 	case CTL_TAG_UNTAGGED:
989 	case CTL_TAG_SIMPLE:
990 	case CTL_TAG_ACA:
991 	default:
992 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
993 		break;
994 	}
995 
996 	if (lbalen->flags & SWS_UNMAP) {
997 		beio->io_offset = lbalen->lba * be_lun->blocksize;
998 		beio->io_len = (uint64_t)lbalen->len * be_lun->blocksize;
999 		beio->bio_cmd = BIO_DELETE;
1000 		beio->ds_trans_type = DEVSTAT_FREE;
1001 
1002 		be_lun->unmap(be_lun, beio);
1003 		return;
1004 	}
1005 
1006 	beio->bio_cmd = BIO_WRITE;
1007 	beio->ds_trans_type = DEVSTAT_WRITE;
1008 
1009 	DPRINTF("WRITE SAME at LBA %jx len %u\n",
1010 	       (uintmax_t)lbalen->lba, lbalen->len);
1011 
1012 	len_left = (uint64_t)lbalen->len * be_lun->blocksize;
1013 	for (i = 0, lba = 0; i < CTLBLK_MAX_SEGS && len_left > 0; i++) {
1014 
1015 		/*
1016 		 * Setup the S/G entry for this chunk.
1017 		 */
1018 		seglen = MIN(CTLBLK_MAX_SEG, len_left);
1019 		seglen -= seglen % be_lun->blocksize;
1020 		beio->sg_segs[i].len = seglen;
1021 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1022 
1023 		DPRINTF("segment %d addr %p len %zd\n", i,
1024 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1025 
1026 		beio->num_segs++;
1027 		len_left -= seglen;
1028 
1029 		buf = beio->sg_segs[i].addr;
1030 		end = buf + seglen;
1031 		for (; buf < end; buf += be_lun->blocksize) {
1032 			memcpy(buf, io->scsiio.kern_data_ptr, be_lun->blocksize);
1033 			if (lbalen->flags & SWS_LBDATA)
1034 				scsi_ulto4b(lbalen->lba + lba, buf);
1035 			lba++;
1036 		}
1037 	}
1038 
1039 	beio->io_offset = lbalen->lba * be_lun->blocksize;
1040 	beio->io_len = lba * be_lun->blocksize;
1041 
1042 	/* We can not do all in one run. Correct and schedule rerun. */
1043 	if (len_left > 0) {
1044 		lbalen->lba += lba;
1045 		lbalen->len -= lba;
1046 		beio->beio_cont = ctl_be_block_cw_done_ws;
1047 	}
1048 
1049 	be_lun->dispatch(be_lun, beio);
1050 }
1051 
1052 static void
1053 ctl_be_block_cw_dispatch_unmap(struct ctl_be_block_lun *be_lun,
1054 			    union ctl_io *io)
1055 {
1056 	struct ctl_be_block_io *beio;
1057 	struct ctl_be_block_softc *softc;
1058 	struct ctl_ptr_len_flags *ptrlen;
1059 
1060 	DPRINTF("entered\n");
1061 
1062 	beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1063 	softc = be_lun->softc;
1064 	ptrlen = (struct ctl_ptr_len_flags *)&io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN];
1065 
1066 	if (ptrlen->flags != 0 || be_lun->unmap == NULL) {
1067 		ctl_free_beio(beio);
1068 		ctl_set_invalid_field(&io->scsiio,
1069 				      /*sks_valid*/ 0,
1070 				      /*command*/ 1,
1071 				      /*field*/ 0,
1072 				      /*bit_valid*/ 0,
1073 				      /*bit*/ 0);
1074 		ctl_config_write_done(io);
1075 		return;
1076 	}
1077 
1078 	/*
1079 	 * If the I/O came down with an ordered or head of queue tag, set
1080 	 * the BIO_ORDERED attribute.  For head of queue tags, that's
1081 	 * pretty much the best we can do.
1082 	 */
1083 	if ((io->scsiio.tag_type == CTL_TAG_ORDERED)
1084 	 || (io->scsiio.tag_type == CTL_TAG_HEAD_OF_QUEUE))
1085 		beio->bio_flags = BIO_ORDERED;
1086 
1087 	switch (io->scsiio.tag_type) {
1088 	case CTL_TAG_ORDERED:
1089 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1090 		break;
1091 	case CTL_TAG_HEAD_OF_QUEUE:
1092 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1093 		break;
1094 	case CTL_TAG_UNTAGGED:
1095 	case CTL_TAG_SIMPLE:
1096 	case CTL_TAG_ACA:
1097 	default:
1098 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1099 		break;
1100 	}
1101 
1102 	beio->io_len = 0;
1103 	beio->io_offset = -1;
1104 
1105 	beio->bio_cmd = BIO_DELETE;
1106 	beio->ds_trans_type = DEVSTAT_FREE;
1107 
1108 	DPRINTF("UNMAP\n");
1109 
1110 	be_lun->unmap(be_lun, beio);
1111 }
1112 
1113 static void
1114 ctl_be_block_cw_done(struct ctl_be_block_io *beio)
1115 {
1116 	union ctl_io *io;
1117 
1118 	io = beio->io;
1119 	ctl_free_beio(beio);
1120 	ctl_config_write_done(io);
1121 }
1122 
1123 static void
1124 ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
1125 			 union ctl_io *io)
1126 {
1127 	struct ctl_be_block_io *beio;
1128 	struct ctl_be_block_softc *softc;
1129 
1130 	DPRINTF("entered\n");
1131 
1132 	softc = be_lun->softc;
1133 	beio = ctl_alloc_beio(softc);
1134 	beio->io = io;
1135 	beio->lun = be_lun;
1136 	beio->beio_cont = ctl_be_block_cw_done;
1137 	PRIV(io)->ptr = (void *)beio;
1138 
1139 	switch (io->scsiio.cdb[0]) {
1140 	case SYNCHRONIZE_CACHE:
1141 	case SYNCHRONIZE_CACHE_16:
1142 		beio->bio_cmd = BIO_FLUSH;
1143 		beio->ds_trans_type = DEVSTAT_NO_DATA;
1144 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1145 		beio->io_len = 0;
1146 		be_lun->lun_flush(be_lun, beio);
1147 		break;
1148 	case WRITE_SAME_10:
1149 	case WRITE_SAME_16:
1150 		ctl_be_block_cw_dispatch_ws(be_lun, io);
1151 		break;
1152 	case UNMAP:
1153 		ctl_be_block_cw_dispatch_unmap(be_lun, io);
1154 		break;
1155 	default:
1156 		panic("Unhandled CDB type %#x", io->scsiio.cdb[0]);
1157 		break;
1158 	}
1159 }
1160 
1161 SDT_PROBE_DEFINE1(cbb, kernel, read, start, "uint64_t");
1162 SDT_PROBE_DEFINE1(cbb, kernel, write, start, "uint64_t");
1163 SDT_PROBE_DEFINE1(cbb, kernel, read, alloc_done, "uint64_t");
1164 SDT_PROBE_DEFINE1(cbb, kernel, write, alloc_done, "uint64_t");
1165 
1166 static void
1167 ctl_be_block_next(struct ctl_be_block_io *beio)
1168 {
1169 	struct ctl_be_block_lun *be_lun;
1170 	union ctl_io *io;
1171 
1172 	io = beio->io;
1173 	be_lun = beio->lun;
1174 	ctl_free_beio(beio);
1175 	if ((io->io_hdr.flags & CTL_FLAG_ABORT) ||
1176 	    ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE &&
1177 	     (io->io_hdr.status & CTL_STATUS_MASK) != CTL_SUCCESS)) {
1178 		ctl_data_submit_done(io);
1179 		return;
1180 	}
1181 
1182 	io->io_hdr.status &= ~CTL_STATUS_MASK;
1183 	io->io_hdr.status |= CTL_STATUS_NONE;
1184 
1185 	mtx_lock(&be_lun->queue_lock);
1186 	/*
1187 	 * XXX KDM make sure that links is okay to use at this point.
1188 	 * Otherwise, we either need to add another field to ctl_io_hdr,
1189 	 * or deal with resource allocation here.
1190 	 */
1191 	STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1192 	mtx_unlock(&be_lun->queue_lock);
1193 
1194 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1195 }
1196 
1197 static void
1198 ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
1199 			   union ctl_io *io)
1200 {
1201 	struct ctl_be_block_io *beio;
1202 	struct ctl_be_block_softc *softc;
1203 	struct ctl_lba_len_flags *lbalen;
1204 	struct ctl_ptr_len_flags *bptrlen;
1205 	uint64_t len_left, lbas;
1206 	int i;
1207 
1208 	softc = be_lun->softc;
1209 
1210 	DPRINTF("entered\n");
1211 
1212 	lbalen = ARGS(io);
1213 	if (lbalen->flags & CTL_LLF_WRITE) {
1214 		SDT_PROBE(cbb, kernel, write, start, 0, 0, 0, 0, 0);
1215 	} else {
1216 		SDT_PROBE(cbb, kernel, read, start, 0, 0, 0, 0, 0);
1217 	}
1218 
1219 	beio = ctl_alloc_beio(softc);
1220 	beio->io = io;
1221 	beio->lun = be_lun;
1222 	bptrlen = PRIV(io);
1223 	bptrlen->ptr = (void *)beio;
1224 
1225 	/*
1226 	 * If the I/O came down with an ordered or head of queue tag, set
1227 	 * the BIO_ORDERED attribute.  For head of queue tags, that's
1228 	 * pretty much the best we can do.
1229 	 *
1230 	 * XXX KDM we don't have a great way to easily know about the FUA
1231 	 * bit right now (it is decoded in ctl_read_write(), but we don't
1232 	 * pass that knowledge to the backend), and in any case we would
1233 	 * need to determine how to handle it.
1234 	 */
1235 	if ((io->scsiio.tag_type == CTL_TAG_ORDERED)
1236 	 || (io->scsiio.tag_type == CTL_TAG_HEAD_OF_QUEUE))
1237 		beio->bio_flags = BIO_ORDERED;
1238 
1239 	switch (io->scsiio.tag_type) {
1240 	case CTL_TAG_ORDERED:
1241 		beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1242 		break;
1243 	case CTL_TAG_HEAD_OF_QUEUE:
1244 		beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1245 		break;
1246 	case CTL_TAG_UNTAGGED:
1247 	case CTL_TAG_SIMPLE:
1248 	case CTL_TAG_ACA:
1249 	default:
1250 		beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1251 		break;
1252 	}
1253 
1254 	if (lbalen->flags & CTL_LLF_WRITE) {
1255 		beio->bio_cmd = BIO_WRITE;
1256 		beio->ds_trans_type = DEVSTAT_WRITE;
1257 	} else {
1258 		beio->bio_cmd = BIO_READ;
1259 		beio->ds_trans_type = DEVSTAT_READ;
1260 	}
1261 
1262 	DPRINTF("%s at LBA %jx len %u @%ju\n",
1263 	       (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE",
1264 	       (uintmax_t)lbalen->lba, lbalen->len, bptrlen->len);
1265 	if (lbalen->flags & CTL_LLF_COMPARE)
1266 		lbas = CTLBLK_HALF_IO_SIZE;
1267 	else
1268 		lbas = CTLBLK_MAX_IO_SIZE;
1269 	lbas = MIN(lbalen->len - bptrlen->len, lbas / be_lun->blocksize);
1270 	beio->io_offset = (lbalen->lba + bptrlen->len) * be_lun->blocksize;
1271 	beio->io_len = lbas * be_lun->blocksize;
1272 	bptrlen->len += lbas;
1273 
1274 	for (i = 0, len_left = beio->io_len; len_left > 0; i++) {
1275 		KASSERT(i < CTLBLK_MAX_SEGS, ("Too many segs (%d >= %d)",
1276 		    i, CTLBLK_MAX_SEGS));
1277 
1278 		/*
1279 		 * Setup the S/G entry for this chunk.
1280 		 */
1281 		beio->sg_segs[i].len = min(CTLBLK_MAX_SEG, len_left);
1282 		beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1283 
1284 		DPRINTF("segment %d addr %p len %zd\n", i,
1285 			beio->sg_segs[i].addr, beio->sg_segs[i].len);
1286 
1287 		/* Set up second segment for compare operation. */
1288 		if (lbalen->flags & CTL_LLF_COMPARE) {
1289 			beio->sg_segs[i + CTLBLK_HALF_SEGS].len =
1290 			    beio->sg_segs[i].len;
1291 			beio->sg_segs[i + CTLBLK_HALF_SEGS].addr =
1292 			    uma_zalloc(be_lun->lun_zone, M_WAITOK);
1293 		}
1294 
1295 		beio->num_segs++;
1296 		len_left -= beio->sg_segs[i].len;
1297 	}
1298 	if (bptrlen->len < lbalen->len)
1299 		beio->beio_cont = ctl_be_block_next;
1300 	io->scsiio.be_move_done = ctl_be_block_move_done;
1301 	/* For compare we have separate S/G lists for read and datamove. */
1302 	if (lbalen->flags & CTL_LLF_COMPARE)
1303 		io->scsiio.kern_data_ptr = (uint8_t *)&beio->sg_segs[CTLBLK_HALF_SEGS];
1304 	else
1305 		io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
1306 	io->scsiio.kern_data_len = beio->io_len;
1307 	io->scsiio.kern_data_resid = 0;
1308 	io->scsiio.kern_sg_entries = beio->num_segs;
1309 	io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
1310 
1311 	/*
1312 	 * For the read case, we need to read the data into our buffers and
1313 	 * then we can send it back to the user.  For the write case, we
1314 	 * need to get the data from the user first.
1315 	 */
1316 	if (beio->bio_cmd == BIO_READ) {
1317 		SDT_PROBE(cbb, kernel, read, alloc_done, 0, 0, 0, 0, 0);
1318 		be_lun->dispatch(be_lun, beio);
1319 	} else {
1320 		SDT_PROBE(cbb, kernel, write, alloc_done, 0, 0, 0, 0, 0);
1321 #ifdef CTL_TIME_IO
1322         	getbintime(&io->io_hdr.dma_start_bt);
1323 #endif
1324 		ctl_datamove(io);
1325 	}
1326 }
1327 
1328 static void
1329 ctl_be_block_worker(void *context, int pending)
1330 {
1331 	struct ctl_be_block_lun *be_lun;
1332 	struct ctl_be_block_softc *softc;
1333 	union ctl_io *io;
1334 
1335 	be_lun = (struct ctl_be_block_lun *)context;
1336 	softc = be_lun->softc;
1337 
1338 	DPRINTF("entered\n");
1339 
1340 	mtx_lock(&be_lun->queue_lock);
1341 	for (;;) {
1342 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->datamove_queue);
1343 		if (io != NULL) {
1344 			struct ctl_be_block_io *beio;
1345 
1346 			DPRINTF("datamove queue\n");
1347 
1348 			STAILQ_REMOVE(&be_lun->datamove_queue, &io->io_hdr,
1349 				      ctl_io_hdr, links);
1350 
1351 			mtx_unlock(&be_lun->queue_lock);
1352 
1353 			beio = (struct ctl_be_block_io *)PRIV(io)->ptr;
1354 
1355 			be_lun->dispatch(be_lun, beio);
1356 
1357 			mtx_lock(&be_lun->queue_lock);
1358 			continue;
1359 		}
1360 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->config_write_queue);
1361 		if (io != NULL) {
1362 
1363 			DPRINTF("config write queue\n");
1364 
1365 			STAILQ_REMOVE(&be_lun->config_write_queue, &io->io_hdr,
1366 				      ctl_io_hdr, links);
1367 
1368 			mtx_unlock(&be_lun->queue_lock);
1369 
1370 			ctl_be_block_cw_dispatch(be_lun, io);
1371 
1372 			mtx_lock(&be_lun->queue_lock);
1373 			continue;
1374 		}
1375 		io = (union ctl_io *)STAILQ_FIRST(&be_lun->input_queue);
1376 		if (io != NULL) {
1377 			DPRINTF("input queue\n");
1378 
1379 			STAILQ_REMOVE(&be_lun->input_queue, &io->io_hdr,
1380 				      ctl_io_hdr, links);
1381 			mtx_unlock(&be_lun->queue_lock);
1382 
1383 			/*
1384 			 * We must drop the lock, since this routine and
1385 			 * its children may sleep.
1386 			 */
1387 			ctl_be_block_dispatch(be_lun, io);
1388 
1389 			mtx_lock(&be_lun->queue_lock);
1390 			continue;
1391 		}
1392 
1393 		/*
1394 		 * If we get here, there is no work left in the queues, so
1395 		 * just break out and let the task queue go to sleep.
1396 		 */
1397 		break;
1398 	}
1399 	mtx_unlock(&be_lun->queue_lock);
1400 }
1401 
1402 /*
1403  * Entry point from CTL to the backend for I/O.  We queue everything to a
1404  * work thread, so this just puts the I/O on a queue and wakes up the
1405  * thread.
1406  */
1407 static int
1408 ctl_be_block_submit(union ctl_io *io)
1409 {
1410 	struct ctl_be_block_lun *be_lun;
1411 	struct ctl_be_lun *ctl_be_lun;
1412 
1413 	DPRINTF("entered\n");
1414 
1415 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
1416 		CTL_PRIV_BACKEND_LUN].ptr;
1417 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
1418 
1419 	/*
1420 	 * Make sure we only get SCSI I/O.
1421 	 */
1422 	KASSERT(io->io_hdr.io_type == CTL_IO_SCSI, ("Non-SCSI I/O (type "
1423 		"%#x) encountered", io->io_hdr.io_type));
1424 
1425 	PRIV(io)->len = 0;
1426 
1427 	mtx_lock(&be_lun->queue_lock);
1428 	/*
1429 	 * XXX KDM make sure that links is okay to use at this point.
1430 	 * Otherwise, we either need to add another field to ctl_io_hdr,
1431 	 * or deal with resource allocation here.
1432 	 */
1433 	STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1434 	mtx_unlock(&be_lun->queue_lock);
1435 	taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1436 
1437 	return (CTL_RETVAL_COMPLETE);
1438 }
1439 
1440 static int
1441 ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
1442 			int flag, struct thread *td)
1443 {
1444 	struct ctl_be_block_softc *softc;
1445 	int error;
1446 
1447 	softc = &backend_block_softc;
1448 
1449 	error = 0;
1450 
1451 	switch (cmd) {
1452 	case CTL_LUN_REQ: {
1453 		struct ctl_lun_req *lun_req;
1454 
1455 		lun_req = (struct ctl_lun_req *)addr;
1456 
1457 		switch (lun_req->reqtype) {
1458 		case CTL_LUNREQ_CREATE:
1459 			error = ctl_be_block_create(softc, lun_req);
1460 			break;
1461 		case CTL_LUNREQ_RM:
1462 			error = ctl_be_block_rm(softc, lun_req);
1463 			break;
1464 		case CTL_LUNREQ_MODIFY:
1465 			error = ctl_be_block_modify(softc, lun_req);
1466 			break;
1467 		default:
1468 			lun_req->status = CTL_LUN_ERROR;
1469 			snprintf(lun_req->error_str, sizeof(lun_req->error_str),
1470 				 "%s: invalid LUN request type %d", __func__,
1471 				 lun_req->reqtype);
1472 			break;
1473 		}
1474 		break;
1475 	}
1476 	default:
1477 		error = ENOTTY;
1478 		break;
1479 	}
1480 
1481 	return (error);
1482 }
1483 
1484 static int
1485 ctl_be_block_open_file(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1486 {
1487 	struct ctl_be_block_filedata *file_data;
1488 	struct ctl_lun_create_params *params;
1489 	struct vattr		      vattr;
1490 	int			      error;
1491 
1492 	error = 0;
1493 	file_data = &be_lun->backend.file;
1494 	params = &req->reqdata.create;
1495 
1496 	be_lun->dev_type = CTL_BE_BLOCK_FILE;
1497 	be_lun->dispatch = ctl_be_block_dispatch_file;
1498 	be_lun->lun_flush = ctl_be_block_flush_file;
1499 
1500 	error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
1501 	if (error != 0) {
1502 		snprintf(req->error_str, sizeof(req->error_str),
1503 			 "error calling VOP_GETATTR() for file %s",
1504 			 be_lun->dev_path);
1505 		return (error);
1506 	}
1507 
1508 	/*
1509 	 * Verify that we have the ability to upgrade to exclusive
1510 	 * access on this file so we can trap errors at open instead
1511 	 * of reporting them during first access.
1512 	 */
1513 	if (VOP_ISLOCKED(be_lun->vn) != LK_EXCLUSIVE) {
1514 		vn_lock(be_lun->vn, LK_UPGRADE | LK_RETRY);
1515 		if (be_lun->vn->v_iflag & VI_DOOMED) {
1516 			error = EBADF;
1517 			snprintf(req->error_str, sizeof(req->error_str),
1518 				 "error locking file %s", be_lun->dev_path);
1519 			return (error);
1520 		}
1521 	}
1522 
1523 
1524 	file_data->cred = crhold(curthread->td_ucred);
1525 	if (params->lun_size_bytes != 0)
1526 		be_lun->size_bytes = params->lun_size_bytes;
1527 	else
1528 		be_lun->size_bytes = vattr.va_size;
1529 	/*
1530 	 * We set the multi thread flag for file operations because all
1531 	 * filesystems (in theory) are capable of allowing multiple readers
1532 	 * of a file at once.  So we want to get the maximum possible
1533 	 * concurrency.
1534 	 */
1535 	be_lun->flags |= CTL_BE_BLOCK_LUN_MULTI_THREAD;
1536 
1537 	/*
1538 	 * XXX KDM vattr.va_blocksize may be larger than 512 bytes here.
1539 	 * With ZFS, it is 131072 bytes.  Block sizes that large don't work
1540 	 * with disklabel and UFS on FreeBSD at least.  Large block sizes
1541 	 * may not work with other OSes as well.  So just export a sector
1542 	 * size of 512 bytes, which should work with any OS or
1543 	 * application.  Since our backing is a file, any block size will
1544 	 * work fine for the backing store.
1545 	 */
1546 #if 0
1547 	be_lun->blocksize= vattr.va_blocksize;
1548 #endif
1549 	if (params->blocksize_bytes != 0)
1550 		be_lun->blocksize = params->blocksize_bytes;
1551 	else
1552 		be_lun->blocksize = 512;
1553 
1554 	/*
1555 	 * Sanity check.  The media size has to be at least one
1556 	 * sector long.
1557 	 */
1558 	if (be_lun->size_bytes < be_lun->blocksize) {
1559 		error = EINVAL;
1560 		snprintf(req->error_str, sizeof(req->error_str),
1561 			 "file %s size %ju < block size %u", be_lun->dev_path,
1562 			 (uintmax_t)be_lun->size_bytes, be_lun->blocksize);
1563 	}
1564 	return (error);
1565 }
1566 
1567 static int
1568 ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1569 {
1570 	struct ctl_lun_create_params *params;
1571 	struct vattr		      vattr;
1572 	struct cdev		     *dev;
1573 	struct cdevsw		     *devsw;
1574 	int			      error;
1575 	off_t			      ps, pss, po, pos;
1576 
1577 	params = &req->reqdata.create;
1578 
1579 	be_lun->dev_type = CTL_BE_BLOCK_DEV;
1580 	be_lun->dispatch = ctl_be_block_dispatch_dev;
1581 	be_lun->lun_flush = ctl_be_block_flush_dev;
1582 	be_lun->unmap = ctl_be_block_unmap_dev;
1583 	be_lun->backend.dev.cdev = be_lun->vn->v_rdev;
1584 	be_lun->backend.dev.csw = dev_refthread(be_lun->backend.dev.cdev,
1585 					     &be_lun->backend.dev.dev_ref);
1586 	if (be_lun->backend.dev.csw == NULL)
1587 		panic("Unable to retrieve device switch");
1588 
1589 	error = VOP_GETATTR(be_lun->vn, &vattr, NOCRED);
1590 	if (error) {
1591 		snprintf(req->error_str, sizeof(req->error_str),
1592 			 "%s: error getting vnode attributes for device %s",
1593 			 __func__, be_lun->dev_path);
1594 		return (error);
1595 	}
1596 
1597 	dev = be_lun->vn->v_rdev;
1598 	devsw = dev->si_devsw;
1599 	if (!devsw->d_ioctl) {
1600 		snprintf(req->error_str, sizeof(req->error_str),
1601 			 "%s: no d_ioctl for device %s!", __func__,
1602 			 be_lun->dev_path);
1603 		return (ENODEV);
1604 	}
1605 
1606 	error = devsw->d_ioctl(dev, DIOCGSECTORSIZE,
1607 			       (caddr_t)&be_lun->blocksize, FREAD,
1608 			       curthread);
1609 	if (error) {
1610 		snprintf(req->error_str, sizeof(req->error_str),
1611 			 "%s: error %d returned for DIOCGSECTORSIZE ioctl "
1612 			 "on %s!", __func__, error, be_lun->dev_path);
1613 		return (error);
1614 	}
1615 
1616 	/*
1617 	 * If the user has asked for a blocksize that is greater than the
1618 	 * backing device's blocksize, we can do it only if the blocksize
1619 	 * the user is asking for is an even multiple of the underlying
1620 	 * device's blocksize.
1621 	 */
1622 	if ((params->blocksize_bytes != 0)
1623 	 && (params->blocksize_bytes > be_lun->blocksize)) {
1624 		uint32_t bs_multiple, tmp_blocksize;
1625 
1626 		bs_multiple = params->blocksize_bytes / be_lun->blocksize;
1627 
1628 		tmp_blocksize = bs_multiple * be_lun->blocksize;
1629 
1630 		if (tmp_blocksize == params->blocksize_bytes) {
1631 			be_lun->blocksize = params->blocksize_bytes;
1632 		} else {
1633 			snprintf(req->error_str, sizeof(req->error_str),
1634 				 "%s: requested blocksize %u is not an even "
1635 				 "multiple of backing device blocksize %u",
1636 				 __func__, params->blocksize_bytes,
1637 				 be_lun->blocksize);
1638 			return (EINVAL);
1639 
1640 		}
1641 	} else if ((params->blocksize_bytes != 0)
1642 		&& (params->blocksize_bytes != be_lun->blocksize)) {
1643 		snprintf(req->error_str, sizeof(req->error_str),
1644 			 "%s: requested blocksize %u < backing device "
1645 			 "blocksize %u", __func__, params->blocksize_bytes,
1646 			 be_lun->blocksize);
1647 		return (EINVAL);
1648 	}
1649 
1650 	error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
1651 			       (caddr_t)&be_lun->size_bytes, FREAD,
1652 			       curthread);
1653 	if (error) {
1654 		snprintf(req->error_str, sizeof(req->error_str),
1655 			 "%s: error %d returned for DIOCGMEDIASIZE "
1656 			 " ioctl on %s!", __func__, error,
1657 			 be_lun->dev_path);
1658 		return (error);
1659 	}
1660 
1661 	if (params->lun_size_bytes != 0) {
1662 		if (params->lun_size_bytes > be_lun->size_bytes) {
1663 			snprintf(req->error_str, sizeof(req->error_str),
1664 				 "%s: requested LUN size %ju > backing device "
1665 				 "size %ju", __func__,
1666 				 (uintmax_t)params->lun_size_bytes,
1667 				 (uintmax_t)be_lun->size_bytes);
1668 			return (EINVAL);
1669 		}
1670 
1671 		be_lun->size_bytes = params->lun_size_bytes;
1672 	}
1673 
1674 	error = devsw->d_ioctl(dev, DIOCGSTRIPESIZE,
1675 			       (caddr_t)&ps, FREAD, curthread);
1676 	if (error)
1677 		ps = po = 0;
1678 	else {
1679 		error = devsw->d_ioctl(dev, DIOCGSTRIPEOFFSET,
1680 				       (caddr_t)&po, FREAD, curthread);
1681 		if (error)
1682 			po = 0;
1683 	}
1684 	pss = ps / be_lun->blocksize;
1685 	pos = po / be_lun->blocksize;
1686 	if ((pss > 0) && (pss * be_lun->blocksize == ps) && (pss >= pos) &&
1687 	    ((pss & (pss - 1)) == 0) && (pos * be_lun->blocksize == po)) {
1688 		be_lun->pblockexp = fls(pss) - 1;
1689 		be_lun->pblockoff = (pss - pos) % pss;
1690 	}
1691 
1692 	return (0);
1693 }
1694 
1695 static int
1696 ctl_be_block_close(struct ctl_be_block_lun *be_lun)
1697 {
1698 	DROP_GIANT();
1699 	if (be_lun->vn) {
1700 		int flags = FREAD | FWRITE;
1701 
1702 		switch (be_lun->dev_type) {
1703 		case CTL_BE_BLOCK_DEV:
1704 			if (be_lun->backend.dev.csw) {
1705 				dev_relthread(be_lun->backend.dev.cdev,
1706 					      be_lun->backend.dev.dev_ref);
1707 				be_lun->backend.dev.csw  = NULL;
1708 				be_lun->backend.dev.cdev = NULL;
1709 			}
1710 			break;
1711 		case CTL_BE_BLOCK_FILE:
1712 			break;
1713 		case CTL_BE_BLOCK_NONE:
1714 			break;
1715 		default:
1716 			panic("Unexpected backend type.");
1717 			break;
1718 		}
1719 
1720 		(void)vn_close(be_lun->vn, flags, NOCRED, curthread);
1721 		be_lun->vn = NULL;
1722 
1723 		switch (be_lun->dev_type) {
1724 		case CTL_BE_BLOCK_DEV:
1725 			break;
1726 		case CTL_BE_BLOCK_FILE:
1727 			if (be_lun->backend.file.cred != NULL) {
1728 				crfree(be_lun->backend.file.cred);
1729 				be_lun->backend.file.cred = NULL;
1730 			}
1731 			break;
1732 		case CTL_BE_BLOCK_NONE:
1733 			break;
1734 		default:
1735 			panic("Unexpected backend type.");
1736 			break;
1737 		}
1738 	}
1739 	PICKUP_GIANT();
1740 
1741 	return (0);
1742 }
1743 
1744 static int
1745 ctl_be_block_open(struct ctl_be_block_softc *softc,
1746 		       struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1747 {
1748 	struct nameidata nd;
1749 	int		 flags;
1750 	int		 error;
1751 
1752 	/*
1753 	 * XXX KDM allow a read-only option?
1754 	 */
1755 	flags = FREAD | FWRITE;
1756 	error = 0;
1757 
1758 	if (rootvnode == NULL) {
1759 		snprintf(req->error_str, sizeof(req->error_str),
1760 			 "%s: Root filesystem is not mounted", __func__);
1761 		return (1);
1762 	}
1763 
1764 	if (!curthread->td_proc->p_fd->fd_cdir) {
1765 		curthread->td_proc->p_fd->fd_cdir = rootvnode;
1766 		VREF(rootvnode);
1767 	}
1768 	if (!curthread->td_proc->p_fd->fd_rdir) {
1769 		curthread->td_proc->p_fd->fd_rdir = rootvnode;
1770 		VREF(rootvnode);
1771 	}
1772 	if (!curthread->td_proc->p_fd->fd_jdir) {
1773 		curthread->td_proc->p_fd->fd_jdir = rootvnode;
1774 		VREF(rootvnode);
1775 	}
1776 
1777  again:
1778 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, be_lun->dev_path, curthread);
1779 	error = vn_open(&nd, &flags, 0, NULL);
1780 	if (error) {
1781 		/*
1782 		 * This is the only reasonable guess we can make as far as
1783 		 * path if the user doesn't give us a fully qualified path.
1784 		 * If they want to specify a file, they need to specify the
1785 		 * full path.
1786 		 */
1787 		if (be_lun->dev_path[0] != '/') {
1788 			char *dev_path = "/dev/";
1789 			char *dev_name;
1790 
1791 			/* Try adding device path at beginning of name */
1792 			dev_name = malloc(strlen(be_lun->dev_path)
1793 					+ strlen(dev_path) + 1,
1794 					  M_CTLBLK, M_WAITOK);
1795 			if (dev_name) {
1796 				sprintf(dev_name, "%s%s", dev_path,
1797 					be_lun->dev_path);
1798 				free(be_lun->dev_path, M_CTLBLK);
1799 				be_lun->dev_path = dev_name;
1800 				goto again;
1801 			}
1802 		}
1803 		snprintf(req->error_str, sizeof(req->error_str),
1804 			 "%s: error opening %s", __func__, be_lun->dev_path);
1805 		return (error);
1806 	}
1807 
1808 	NDFREE(&nd, NDF_ONLY_PNBUF);
1809 
1810 	be_lun->vn = nd.ni_vp;
1811 
1812 	/* We only support disks and files. */
1813 	if (vn_isdisk(be_lun->vn, &error)) {
1814 		error = ctl_be_block_open_dev(be_lun, req);
1815 	} else if (be_lun->vn->v_type == VREG) {
1816 		error = ctl_be_block_open_file(be_lun, req);
1817 	} else {
1818 		error = EINVAL;
1819 		snprintf(req->error_str, sizeof(req->error_str),
1820 			 "%s is not a disk or plain file", be_lun->dev_path);
1821 	}
1822 	VOP_UNLOCK(be_lun->vn, 0);
1823 
1824 	if (error != 0) {
1825 		ctl_be_block_close(be_lun);
1826 		return (error);
1827 	}
1828 
1829 	be_lun->blocksize_shift = fls(be_lun->blocksize) - 1;
1830 	be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
1831 
1832 	return (0);
1833 }
1834 
1835 static int
1836 ctl_be_block_create(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
1837 {
1838 	struct ctl_be_block_lun *be_lun;
1839 	struct ctl_lun_create_params *params;
1840 	char num_thread_str[16];
1841 	char tmpstr[32];
1842 	char *value;
1843 	int retval, num_threads, unmap;
1844 	int tmp_num_threads;
1845 
1846 	params = &req->reqdata.create;
1847 	retval = 0;
1848 
1849 	num_threads = cbb_num_threads;
1850 
1851 	be_lun = malloc(sizeof(*be_lun), M_CTLBLK, M_ZERO | M_WAITOK);
1852 
1853 	be_lun->softc = softc;
1854 	STAILQ_INIT(&be_lun->input_queue);
1855 	STAILQ_INIT(&be_lun->config_write_queue);
1856 	STAILQ_INIT(&be_lun->datamove_queue);
1857 	sprintf(be_lun->lunname, "cblk%d", softc->num_luns);
1858 	mtx_init(&be_lun->io_lock, "cblk io lock", NULL, MTX_DEF);
1859 	mtx_init(&be_lun->queue_lock, "cblk queue lock", NULL, MTX_DEF);
1860 	ctl_init_opts(&be_lun->ctl_be_lun.options,
1861 	    req->num_be_args, req->kern_be_args);
1862 
1863 	be_lun->lun_zone = uma_zcreate(be_lun->lunname, CTLBLK_MAX_SEG,
1864 	    NULL, NULL, NULL, NULL, /*align*/ 0, /*flags*/0);
1865 
1866 	if (be_lun->lun_zone == NULL) {
1867 		snprintf(req->error_str, sizeof(req->error_str),
1868 			 "%s: error allocating UMA zone", __func__);
1869 		goto bailout_error;
1870 	}
1871 
1872 	if (params->flags & CTL_LUN_FLAG_DEV_TYPE)
1873 		be_lun->ctl_be_lun.lun_type = params->device_type;
1874 	else
1875 		be_lun->ctl_be_lun.lun_type = T_DIRECT;
1876 
1877 	if (be_lun->ctl_be_lun.lun_type == T_DIRECT) {
1878 		value = ctl_get_opt(&be_lun->ctl_be_lun.options, "file");
1879 		if (value == NULL) {
1880 			snprintf(req->error_str, sizeof(req->error_str),
1881 				 "%s: no file argument specified", __func__);
1882 			goto bailout_error;
1883 		}
1884 		be_lun->dev_path = strdup(value, M_CTLBLK);
1885 
1886 		retval = ctl_be_block_open(softc, be_lun, req);
1887 		if (retval != 0) {
1888 			retval = 0;
1889 			goto bailout_error;
1890 		}
1891 
1892 		/*
1893 		 * Tell the user the size of the file/device.
1894 		 */
1895 		params->lun_size_bytes = be_lun->size_bytes;
1896 
1897 		/*
1898 		 * The maximum LBA is the size - 1.
1899 		 */
1900 		be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
1901 	} else {
1902 		/*
1903 		 * For processor devices, we don't have any size.
1904 		 */
1905 		be_lun->blocksize = 0;
1906 		be_lun->pblockexp = 0;
1907 		be_lun->pblockoff = 0;
1908 		be_lun->size_blocks = 0;
1909 		be_lun->size_bytes = 0;
1910 		be_lun->ctl_be_lun.maxlba = 0;
1911 		params->lun_size_bytes = 0;
1912 
1913 		/*
1914 		 * Default to just 1 thread for processor devices.
1915 		 */
1916 		num_threads = 1;
1917 	}
1918 
1919 	/*
1920 	 * XXX This searching loop might be refactored to be combined with
1921 	 * the loop above,
1922 	 */
1923 	value = ctl_get_opt(&be_lun->ctl_be_lun.options, "num_threads");
1924 	if (value != NULL) {
1925 		tmp_num_threads = strtol(value, NULL, 0);
1926 
1927 		/*
1928 		 * We don't let the user specify less than one
1929 		 * thread, but hope he's clueful enough not to
1930 		 * specify 1000 threads.
1931 		 */
1932 		if (tmp_num_threads < 1) {
1933 			snprintf(req->error_str, sizeof(req->error_str),
1934 				 "%s: invalid number of threads %s",
1935 			         __func__, num_thread_str);
1936 			goto bailout_error;
1937 		}
1938 		num_threads = tmp_num_threads;
1939 	}
1940 	unmap = 0;
1941 	value = ctl_get_opt(&be_lun->ctl_be_lun.options, "unmap");
1942 	if (value != NULL && strcmp(value, "on") == 0)
1943 		unmap = 1;
1944 
1945 	be_lun->flags = CTL_BE_BLOCK_LUN_UNCONFIGURED;
1946 	be_lun->ctl_be_lun.flags = CTL_LUN_FLAG_PRIMARY;
1947 	if (unmap)
1948 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_UNMAP;
1949 	be_lun->ctl_be_lun.be_lun = be_lun;
1950 	be_lun->ctl_be_lun.blocksize = be_lun->blocksize;
1951 	be_lun->ctl_be_lun.pblockexp = be_lun->pblockexp;
1952 	be_lun->ctl_be_lun.pblockoff = be_lun->pblockoff;
1953 	/* Tell the user the blocksize we ended up using */
1954 	params->blocksize_bytes = be_lun->blocksize;
1955 	if (params->flags & CTL_LUN_FLAG_ID_REQ) {
1956 		be_lun->ctl_be_lun.req_lun_id = params->req_lun_id;
1957 		be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_ID_REQ;
1958 	} else
1959 		be_lun->ctl_be_lun.req_lun_id = 0;
1960 
1961 	be_lun->ctl_be_lun.lun_shutdown = ctl_be_block_lun_shutdown;
1962 	be_lun->ctl_be_lun.lun_config_status =
1963 		ctl_be_block_lun_config_status;
1964 	be_lun->ctl_be_lun.be = &ctl_be_block_driver;
1965 
1966 	if ((params->flags & CTL_LUN_FLAG_SERIAL_NUM) == 0) {
1967 		snprintf(tmpstr, sizeof(tmpstr), "MYSERIAL%4d",
1968 			 softc->num_luns);
1969 		strncpy((char *)be_lun->ctl_be_lun.serial_num, tmpstr,
1970 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
1971 			sizeof(tmpstr)));
1972 
1973 		/* Tell the user what we used for a serial number */
1974 		strncpy((char *)params->serial_num, tmpstr,
1975 			ctl_min(sizeof(params->serial_num), sizeof(tmpstr)));
1976 	} else {
1977 		strncpy((char *)be_lun->ctl_be_lun.serial_num,
1978 			params->serial_num,
1979 			ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
1980 			sizeof(params->serial_num)));
1981 	}
1982 	if ((params->flags & CTL_LUN_FLAG_DEVID) == 0) {
1983 		snprintf(tmpstr, sizeof(tmpstr), "MYDEVID%4d", softc->num_luns);
1984 		strncpy((char *)be_lun->ctl_be_lun.device_id, tmpstr,
1985 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
1986 			sizeof(tmpstr)));
1987 
1988 		/* Tell the user what we used for a device ID */
1989 		strncpy((char *)params->device_id, tmpstr,
1990 			ctl_min(sizeof(params->device_id), sizeof(tmpstr)));
1991 	} else {
1992 		strncpy((char *)be_lun->ctl_be_lun.device_id,
1993 			params->device_id,
1994 			ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
1995 				sizeof(params->device_id)));
1996 	}
1997 
1998 	TASK_INIT(&be_lun->io_task, /*priority*/0, ctl_be_block_worker, be_lun);
1999 
2000 	be_lun->io_taskqueue = taskqueue_create(be_lun->lunname, M_WAITOK,
2001 	    taskqueue_thread_enqueue, /*context*/&be_lun->io_taskqueue);
2002 
2003 	if (be_lun->io_taskqueue == NULL) {
2004 		snprintf(req->error_str, sizeof(req->error_str),
2005 			 "%s: Unable to create taskqueue", __func__);
2006 		goto bailout_error;
2007 	}
2008 
2009 	/*
2010 	 * Note that we start the same number of threads by default for
2011 	 * both the file case and the block device case.  For the file
2012 	 * case, we need multiple threads to allow concurrency, because the
2013 	 * vnode interface is designed to be a blocking interface.  For the
2014 	 * block device case, ZFS zvols at least will block the caller's
2015 	 * context in many instances, and so we need multiple threads to
2016 	 * overcome that problem.  Other block devices don't need as many
2017 	 * threads, but they shouldn't cause too many problems.
2018 	 *
2019 	 * If the user wants to just have a single thread for a block
2020 	 * device, he can specify that when the LUN is created, or change
2021 	 * the tunable/sysctl to alter the default number of threads.
2022 	 */
2023 	retval = taskqueue_start_threads(&be_lun->io_taskqueue,
2024 					 /*num threads*/num_threads,
2025 					 /*priority*/PWAIT,
2026 					 /*thread name*/
2027 					 "%s taskq", be_lun->lunname);
2028 
2029 	if (retval != 0)
2030 		goto bailout_error;
2031 
2032 	be_lun->num_threads = num_threads;
2033 
2034 	mtx_lock(&softc->lock);
2035 	softc->num_luns++;
2036 	STAILQ_INSERT_TAIL(&softc->lun_list, be_lun, links);
2037 
2038 	mtx_unlock(&softc->lock);
2039 
2040 	retval = ctl_add_lun(&be_lun->ctl_be_lun);
2041 	if (retval != 0) {
2042 		mtx_lock(&softc->lock);
2043 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2044 			      links);
2045 		softc->num_luns--;
2046 		mtx_unlock(&softc->lock);
2047 		snprintf(req->error_str, sizeof(req->error_str),
2048 			 "%s: ctl_add_lun() returned error %d, see dmesg for "
2049 			"details", __func__, retval);
2050 		retval = 0;
2051 		goto bailout_error;
2052 	}
2053 
2054 	mtx_lock(&softc->lock);
2055 
2056 	/*
2057 	 * Tell the config_status routine that we're waiting so it won't
2058 	 * clean up the LUN in the event of an error.
2059 	 */
2060 	be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
2061 
2062 	while (be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) {
2063 		retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
2064 		if (retval == EINTR)
2065 			break;
2066 	}
2067 	be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
2068 
2069 	if (be_lun->flags & CTL_BE_BLOCK_LUN_CONFIG_ERR) {
2070 		snprintf(req->error_str, sizeof(req->error_str),
2071 			 "%s: LUN configuration error, see dmesg for details",
2072 			 __func__);
2073 		STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
2074 			      links);
2075 		softc->num_luns--;
2076 		mtx_unlock(&softc->lock);
2077 		goto bailout_error;
2078 	} else {
2079 		params->req_lun_id = be_lun->ctl_be_lun.lun_id;
2080 	}
2081 
2082 	mtx_unlock(&softc->lock);
2083 
2084 	be_lun->disk_stats = devstat_new_entry("cbb", params->req_lun_id,
2085 					       be_lun->blocksize,
2086 					       DEVSTAT_ALL_SUPPORTED,
2087 					       be_lun->ctl_be_lun.lun_type
2088 					       | DEVSTAT_TYPE_IF_OTHER,
2089 					       DEVSTAT_PRIORITY_OTHER);
2090 
2091 
2092 	req->status = CTL_LUN_OK;
2093 
2094 	return (retval);
2095 
2096 bailout_error:
2097 	req->status = CTL_LUN_ERROR;
2098 
2099 	if (be_lun->io_taskqueue != NULL)
2100 		taskqueue_free(be_lun->io_taskqueue);
2101 	ctl_be_block_close(be_lun);
2102 	if (be_lun->dev_path != NULL)
2103 		free(be_lun->dev_path, M_CTLBLK);
2104 	if (be_lun->lun_zone != NULL)
2105 		uma_zdestroy(be_lun->lun_zone);
2106 	ctl_free_opts(&be_lun->ctl_be_lun.options);
2107 	mtx_destroy(&be_lun->queue_lock);
2108 	mtx_destroy(&be_lun->io_lock);
2109 	free(be_lun, M_CTLBLK);
2110 
2111 	return (retval);
2112 }
2113 
2114 static int
2115 ctl_be_block_rm(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2116 {
2117 	struct ctl_lun_rm_params *params;
2118 	struct ctl_be_block_lun *be_lun;
2119 	int retval;
2120 
2121 	params = &req->reqdata.rm;
2122 
2123 	mtx_lock(&softc->lock);
2124 
2125 	be_lun = NULL;
2126 
2127 	STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2128 		if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2129 			break;
2130 	}
2131 	mtx_unlock(&softc->lock);
2132 
2133 	if (be_lun == NULL) {
2134 		snprintf(req->error_str, sizeof(req->error_str),
2135 			 "%s: LUN %u is not managed by the block backend",
2136 			 __func__, params->lun_id);
2137 		goto bailout_error;
2138 	}
2139 
2140 	retval = ctl_disable_lun(&be_lun->ctl_be_lun);
2141 
2142 	if (retval != 0) {
2143 		snprintf(req->error_str, sizeof(req->error_str),
2144 			 "%s: error %d returned from ctl_disable_lun() for "
2145 			 "LUN %d", __func__, retval, params->lun_id);
2146 		goto bailout_error;
2147 
2148 	}
2149 
2150 	retval = ctl_invalidate_lun(&be_lun->ctl_be_lun);
2151 	if (retval != 0) {
2152 		snprintf(req->error_str, sizeof(req->error_str),
2153 			 "%s: error %d returned from ctl_invalidate_lun() for "
2154 			 "LUN %d", __func__, retval, params->lun_id);
2155 		goto bailout_error;
2156 	}
2157 
2158 	mtx_lock(&softc->lock);
2159 
2160 	be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
2161 
2162 	while ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
2163                 retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
2164                 if (retval == EINTR)
2165                         break;
2166         }
2167 
2168 	be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
2169 
2170 	if ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
2171 		snprintf(req->error_str, sizeof(req->error_str),
2172 			 "%s: interrupted waiting for LUN to be freed",
2173 			 __func__);
2174 		mtx_unlock(&softc->lock);
2175 		goto bailout_error;
2176 	}
2177 
2178 	STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun, links);
2179 
2180 	softc->num_luns--;
2181 	mtx_unlock(&softc->lock);
2182 
2183 	taskqueue_drain(be_lun->io_taskqueue, &be_lun->io_task);
2184 
2185 	taskqueue_free(be_lun->io_taskqueue);
2186 
2187 	ctl_be_block_close(be_lun);
2188 
2189 	if (be_lun->disk_stats != NULL)
2190 		devstat_remove_entry(be_lun->disk_stats);
2191 
2192 	uma_zdestroy(be_lun->lun_zone);
2193 
2194 	ctl_free_opts(&be_lun->ctl_be_lun.options);
2195 	free(be_lun->dev_path, M_CTLBLK);
2196 	mtx_destroy(&be_lun->queue_lock);
2197 	mtx_destroy(&be_lun->io_lock);
2198 	free(be_lun, M_CTLBLK);
2199 
2200 	req->status = CTL_LUN_OK;
2201 
2202 	return (0);
2203 
2204 bailout_error:
2205 
2206 	req->status = CTL_LUN_ERROR;
2207 
2208 	return (0);
2209 }
2210 
2211 static int
2212 ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
2213 			 struct ctl_lun_req *req)
2214 {
2215 	struct vattr vattr;
2216 	int error;
2217 	struct ctl_lun_modify_params *params;
2218 
2219 	params = &req->reqdata.modify;
2220 
2221 	if (params->lun_size_bytes != 0) {
2222 		be_lun->size_bytes = params->lun_size_bytes;
2223 	} else  {
2224 		error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
2225 		if (error != 0) {
2226 			snprintf(req->error_str, sizeof(req->error_str),
2227 				 "error calling VOP_GETATTR() for file %s",
2228 				 be_lun->dev_path);
2229 			return (error);
2230 		}
2231 
2232 		be_lun->size_bytes = vattr.va_size;
2233 	}
2234 
2235 	return (0);
2236 }
2237 
2238 static int
2239 ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
2240 			struct ctl_lun_req *req)
2241 {
2242 	struct cdev *dev;
2243 	struct cdevsw *devsw;
2244 	int error;
2245 	struct ctl_lun_modify_params *params;
2246 	uint64_t size_bytes;
2247 
2248 	params = &req->reqdata.modify;
2249 
2250 	dev = be_lun->vn->v_rdev;
2251 	devsw = dev->si_devsw;
2252 	if (!devsw->d_ioctl) {
2253 		snprintf(req->error_str, sizeof(req->error_str),
2254 			 "%s: no d_ioctl for device %s!", __func__,
2255 			 be_lun->dev_path);
2256 		return (ENODEV);
2257 	}
2258 
2259 	error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
2260 			       (caddr_t)&size_bytes, FREAD,
2261 			       curthread);
2262 	if (error) {
2263 		snprintf(req->error_str, sizeof(req->error_str),
2264 			 "%s: error %d returned for DIOCGMEDIASIZE ioctl "
2265 			 "on %s!", __func__, error, be_lun->dev_path);
2266 		return (error);
2267 	}
2268 
2269 	if (params->lun_size_bytes != 0) {
2270 		if (params->lun_size_bytes > size_bytes) {
2271 			snprintf(req->error_str, sizeof(req->error_str),
2272 				 "%s: requested LUN size %ju > backing device "
2273 				 "size %ju", __func__,
2274 				 (uintmax_t)params->lun_size_bytes,
2275 				 (uintmax_t)size_bytes);
2276 			return (EINVAL);
2277 		}
2278 
2279 		be_lun->size_bytes = params->lun_size_bytes;
2280 	} else {
2281 		be_lun->size_bytes = size_bytes;
2282 	}
2283 
2284 	return (0);
2285 }
2286 
2287 static int
2288 ctl_be_block_modify(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2289 {
2290 	struct ctl_lun_modify_params *params;
2291 	struct ctl_be_block_lun *be_lun;
2292 	int error;
2293 
2294 	params = &req->reqdata.modify;
2295 
2296 	mtx_lock(&softc->lock);
2297 
2298 	be_lun = NULL;
2299 
2300 	STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2301 		if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2302 			break;
2303 	}
2304 	mtx_unlock(&softc->lock);
2305 
2306 	if (be_lun == NULL) {
2307 		snprintf(req->error_str, sizeof(req->error_str),
2308 			 "%s: LUN %u is not managed by the block backend",
2309 			 __func__, params->lun_id);
2310 		goto bailout_error;
2311 	}
2312 
2313 	if (params->lun_size_bytes != 0) {
2314 		if (params->lun_size_bytes < be_lun->blocksize) {
2315 			snprintf(req->error_str, sizeof(req->error_str),
2316 				"%s: LUN size %ju < blocksize %u", __func__,
2317 				params->lun_size_bytes, be_lun->blocksize);
2318 			goto bailout_error;
2319 		}
2320 	}
2321 
2322 	vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
2323 
2324 	if (be_lun->vn->v_type == VREG)
2325 		error = ctl_be_block_modify_file(be_lun, req);
2326 	else
2327 		error = ctl_be_block_modify_dev(be_lun, req);
2328 
2329 	VOP_UNLOCK(be_lun->vn, 0);
2330 
2331 	if (error != 0)
2332 		goto bailout_error;
2333 
2334 	be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
2335 
2336 	/*
2337 	 * The maximum LBA is the size - 1.
2338 	 *
2339 	 * XXX: Note that this field is being updated without locking,
2340 	 * 	which might cause problems on 32-bit architectures.
2341 	 */
2342 	be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
2343 	ctl_lun_capacity_changed(&be_lun->ctl_be_lun);
2344 
2345 	/* Tell the user the exact size we ended up using */
2346 	params->lun_size_bytes = be_lun->size_bytes;
2347 
2348 	req->status = CTL_LUN_OK;
2349 
2350 	return (0);
2351 
2352 bailout_error:
2353 	req->status = CTL_LUN_ERROR;
2354 
2355 	return (0);
2356 }
2357 
2358 static void
2359 ctl_be_block_lun_shutdown(void *be_lun)
2360 {
2361 	struct ctl_be_block_lun *lun;
2362 	struct ctl_be_block_softc *softc;
2363 
2364 	lun = (struct ctl_be_block_lun *)be_lun;
2365 
2366 	softc = lun->softc;
2367 
2368 	mtx_lock(&softc->lock);
2369 	lun->flags |= CTL_BE_BLOCK_LUN_UNCONFIGURED;
2370 	if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2371 		wakeup(lun);
2372 	mtx_unlock(&softc->lock);
2373 
2374 }
2375 
2376 static void
2377 ctl_be_block_lun_config_status(void *be_lun, ctl_lun_config_status status)
2378 {
2379 	struct ctl_be_block_lun *lun;
2380 	struct ctl_be_block_softc *softc;
2381 
2382 	lun = (struct ctl_be_block_lun *)be_lun;
2383 	softc = lun->softc;
2384 
2385 	if (status == CTL_LUN_CONFIG_OK) {
2386 		mtx_lock(&softc->lock);
2387 		lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2388 		if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2389 			wakeup(lun);
2390 		mtx_unlock(&softc->lock);
2391 
2392 		/*
2393 		 * We successfully added the LUN, attempt to enable it.
2394 		 */
2395 		if (ctl_enable_lun(&lun->ctl_be_lun) != 0) {
2396 			printf("%s: ctl_enable_lun() failed!\n", __func__);
2397 			if (ctl_invalidate_lun(&lun->ctl_be_lun) != 0) {
2398 				printf("%s: ctl_invalidate_lun() failed!\n",
2399 				       __func__);
2400 			}
2401 		}
2402 
2403 		return;
2404 	}
2405 
2406 
2407 	mtx_lock(&softc->lock);
2408 	lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2409 	lun->flags |= CTL_BE_BLOCK_LUN_CONFIG_ERR;
2410 	wakeup(lun);
2411 	mtx_unlock(&softc->lock);
2412 }
2413 
2414 
2415 static int
2416 ctl_be_block_config_write(union ctl_io *io)
2417 {
2418 	struct ctl_be_block_lun *be_lun;
2419 	struct ctl_be_lun *ctl_be_lun;
2420 	int retval;
2421 
2422 	retval = 0;
2423 
2424 	DPRINTF("entered\n");
2425 
2426 	ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
2427 		CTL_PRIV_BACKEND_LUN].ptr;
2428 	be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
2429 
2430 	switch (io->scsiio.cdb[0]) {
2431 	case SYNCHRONIZE_CACHE:
2432 	case SYNCHRONIZE_CACHE_16:
2433 	case WRITE_SAME_10:
2434 	case WRITE_SAME_16:
2435 	case UNMAP:
2436 		/*
2437 		 * The upper level CTL code will filter out any CDBs with
2438 		 * the immediate bit set and return the proper error.
2439 		 *
2440 		 * We don't really need to worry about what LBA range the
2441 		 * user asked to be synced out.  When they issue a sync
2442 		 * cache command, we'll sync out the whole thing.
2443 		 */
2444 		mtx_lock(&be_lun->queue_lock);
2445 		STAILQ_INSERT_TAIL(&be_lun->config_write_queue, &io->io_hdr,
2446 				   links);
2447 		mtx_unlock(&be_lun->queue_lock);
2448 		taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
2449 		break;
2450 	case START_STOP_UNIT: {
2451 		struct scsi_start_stop_unit *cdb;
2452 
2453 		cdb = (struct scsi_start_stop_unit *)io->scsiio.cdb;
2454 
2455 		if (cdb->how & SSS_START)
2456 			retval = ctl_start_lun(ctl_be_lun);
2457 		else {
2458 			retval = ctl_stop_lun(ctl_be_lun);
2459 			/*
2460 			 * XXX KDM Copan-specific offline behavior.
2461 			 * Figure out a reasonable way to port this?
2462 			 */
2463 #ifdef NEEDTOPORT
2464 			if ((retval == 0)
2465 			 && (cdb->byte2 & SSS_ONOFFLINE))
2466 				retval = ctl_lun_offline(ctl_be_lun);
2467 #endif
2468 		}
2469 
2470 		/*
2471 		 * In general, the above routines should not fail.  They
2472 		 * just set state for the LUN.  So we've got something
2473 		 * pretty wrong here if we can't start or stop the LUN.
2474 		 */
2475 		if (retval != 0) {
2476 			ctl_set_internal_failure(&io->scsiio,
2477 						 /*sks_valid*/ 1,
2478 						 /*retry_count*/ 0xf051);
2479 			retval = CTL_RETVAL_COMPLETE;
2480 		} else {
2481 			ctl_set_success(&io->scsiio);
2482 		}
2483 		ctl_config_write_done(io);
2484 		break;
2485 	}
2486 	default:
2487 		ctl_set_invalid_opcode(&io->scsiio);
2488 		ctl_config_write_done(io);
2489 		retval = CTL_RETVAL_COMPLETE;
2490 		break;
2491 	}
2492 
2493 	return (retval);
2494 
2495 }
2496 
2497 static int
2498 ctl_be_block_config_read(union ctl_io *io)
2499 {
2500 	return (0);
2501 }
2502 
2503 static int
2504 ctl_be_block_lun_info(void *be_lun, struct sbuf *sb)
2505 {
2506 	struct ctl_be_block_lun *lun;
2507 	int retval;
2508 
2509 	lun = (struct ctl_be_block_lun *)be_lun;
2510 	retval = 0;
2511 
2512 	retval = sbuf_printf(sb, "\t<num_threads>");
2513 
2514 	if (retval != 0)
2515 		goto bailout;
2516 
2517 	retval = sbuf_printf(sb, "%d", lun->num_threads);
2518 
2519 	if (retval != 0)
2520 		goto bailout;
2521 
2522 	retval = sbuf_printf(sb, "</num_threads>\n");
2523 
2524 bailout:
2525 
2526 	return (retval);
2527 }
2528 
2529 int
2530 ctl_be_block_init(void)
2531 {
2532 	struct ctl_be_block_softc *softc;
2533 	int retval;
2534 
2535 	softc = &backend_block_softc;
2536 	retval = 0;
2537 
2538 	mtx_init(&softc->lock, "ctlblock", NULL, MTX_DEF);
2539 	beio_zone = uma_zcreate("beio", sizeof(struct ctl_be_block_io),
2540 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
2541 	STAILQ_INIT(&softc->disk_list);
2542 	STAILQ_INIT(&softc->lun_list);
2543 
2544 	return (retval);
2545 }
2546